BooksAudiobooksComicsSheet Music### Categories

### Categories

Scribd Selects Books

Hand-picked favorites from

our editors

our editors

Scribd Selects Audiobooks

Hand-picked favorites from

our editors

our editors

Scribd Selects Comics

Hand-picked favorites from

our editors

our editors

Scribd Selects Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

P. 1

Control of Color Imaging Systems, Analysis and Design|Views: 719|Likes: 8

Published by Tuli Pan

See more

See less

https://www.scribd.com/doc/95602139/Control-of-Color-Imaging-Systems-Analysis-and-Design

02/08/2015

text

original

- 1.1 INTRODUCTION
- 1.2 PRINTING AND PUBLISHING SYSTEM
- 1.2.1 BUSINESS MANAGEMENT
- 1.2.2 OUTPUT PRODUCTION
- 1.2.3 PROCESS MANAGEMENT
- 1.3 DIGITAL FRONT END
- 1.4 DIGITAL PRINT ENGINE (ELECTROPHOTOGRAPHIC)
- 1.4.1 IMAGE-ON-IMAGE AND TANDEM PRINT ENGINES
- 1.4.2 PARALLEL PRINTING SYSTEMS
- 1.6 PREPRESS-BASED PROCESSING
- 1.7 DFE-BASED PROCESSING
- 1.8 PRINT ENGINE-BASED PROCESSING
- REFERENCES
- 2.1 INTRODUCTION
- 2.2 DIGITAL IMAGE FORMATION AND SYSTEMS
- 2.2.1 POINT SPREAD FUNCTION OF A DEFOCUSED LENS
- 2.2.2 POINT SPREAD FUNCTION OF MOTION BLUR
- 2.2.3 POINT SPREAD FUNCTION OF HUMAN VISUAL SYSTEM
- 2.3 OPTICAL AND MODULATION TRANSFER FUNCTIONS
- 2.4 IMAGE SAMPLING AND QUANTIZATION
- 2.4.1 TWO-DIMENSIONAL SAMPLING THEOREM
- 2.4.2 IMAGE QUANTIZATION
- 2.4.2.1 Uniform Quantization
- 2.4.2.2 Signal-to-Quantization Noise Ratio (SQNR)
- 2.4.2.3 Optimum Minimum Mean-Square Error Quantizer
- 2.4.2.4 Perceptual Quantization
- 2.4.2.5 Vector Quantization
- 2.5 IMAGE TRANSFORM
- 2.5.1 TWO-DIMENSIONAL DISCRETE FOURIER TRANSFORM
- 2.5.2 TWO-DIMENSIONAL DISCRETE COSINE TRANSFORM
- 2.5.3 TWO-DIMENSIONAL HADAMARD TRANSFORM
- 2.5.3.1 Inverse Hadamard Transform
- 2.6 IMAGE FILTERING
- 2.6.1 DESIGN OF 2-D FIR FILTERS
- 2.7 IMAGE RESIZING
- 2.7.1 DEFINITION OF SAMPLING RATE CONVERSION
- 2.7.2 UPSAMPLING BY FACTOR OF P
- 2.7.3 DOWNSAMPLING BY FACTOR OFQ
- 2.7.4 SAMPLING RATE CONVERSION BY A FACTOR OF P Q
- 2.8 IMAGE ENHANCEMENT
- 2.8.1 UNSHARP MASKING
- 2.8.2 IMAGE HISTOGRAM
- 2.8.3 HISTOGRAM EQUALIZATION
- 2.9 IMAGE RESTORATION
- 2.9.1 WIENER FILTER RESTORATION
- 2.10 IMAGE HALFTONING
- 2.10.1 ERROR DIFFUSION ALGORITHM
- PROBLEMS
- 3.1 INTRODUCTION
- 3.2 GENERAL CONTINUOUS-TIME SYSTEM DESCRIPTION
- 3.3 LAPLACE TRANSFORM
- 3.3.1 INVERSE LAPLACE TRANSFORM
- 3.4 GENERAL LINEAR DISCRETE-TIME SYSTEMS
- 3.4.1 SOLUTION OF CONSTANT-COEFFICIENTS DIFFERENCE EQUATIONS
- 3.5 z-TRANSFORM
- 3.5.1 PROPERTIES OF Z-TRANSFORM
- 3.5.2 INVERSE Z-TRANSFORM
- 3.5.3 RELATION BETWEEN THE Z-TRANSFORM AND THE LAPLACE TRANSFORM
- 3.6 DISCRETE-TIME FOURIER TRANSFORM
- 3.6.1 PROPERTIES OF DISCRETE-TIME FOURIER TRANSFORM
- 3.6.2 INVERSE DTFT
- 3.7 TWO-DIMENSIONAL z-TRANSFORM
- 3.8 TWO-DIMENSIONAL DISCRETE-SPACE FOURIER TRANSFORM
- 3.8.1 PROPERTIES OF 2-D DSFT
- 3.8.2 INVERSE 2-D DSFT
- 3.9 EIGENVALUES AND EIGENVECTORS
- 3.9.1 DEFINITION OF EIGENVALUE AND EIGENVECTOR
- 3.9.2 PRODUCT AND SUM OF EIGENVALUES
- 3.9.3 FINDING CHARACTERISTIC POLYNOMIAL OF A MATRIX
- 3.9.4 MODAL MATRIX
- 3.9.5 MATRIX DIAGONALIZATION
- 3.10 SINGULAR VALUE DECOMPOSITION
- 3.10.1 MATRIX NORM
- 3.10.2 PRINCIPAL COMPONENTS ANALYSIS
- 3.11 MATRIX POLYNOMIALS AND FUNCTIONS OF SQUARE MATRICES
- 3.11.1 MATRIX POLYNOMIAL
- 3.11.2 INFINITE SERIES OF MATRICES
- 3.11.3 CAYLEY–HAMILTON THEOREM
- 3.11.4 FUNCTION OF MATRICES
- 3.11.4.1 Cayley–Hamilton Technique
- 3.11.4.2 Modal-Matrix Technique
- 3.11.5 MATRIX EXPONENTIAL FUNCTION eAt
- 3.11.7 MATRIX EXPONENTIAL FUNCTION Ak
- 3.12 FUNDAMENTALS OF MATRIX CALCULUS
- 3.12.1 DERIVATIVES OF A SCALAR FUNCTION WITH RESPECT TO A VECTOR
- 3.12.2 DERIVATIVES OF QUADRATIC FUNCTIONS
- 3.12.3 DERIVATIVE OF A VECTOR FUNCTION WITH RESPECT TO A VECTOR
- 4.1 INTRODUCTION
- 4.2 CONCEPT OF STATES
- 4.3 STATE-SPACE REPRESENTATION OF CONTINUOUS-TIME SYSTEMS
- 4.3.1 DEFINITION OF STATE
- 4.3.2 STATE EQUATIONS OF CONTINUOUS-TIME SYSTEMS
- 4.3.3 STATE-SPACE EQUATIONS OF ELECTRICAL SYSTEMS
- 4.3.4 STATE-SPACE EQUATIONS OF MECHANICAL SYSTEMS
- 4.4 STATE-SPACE REPRESENTATION OF GENERAL CONTINUOUS LTI SYSTEMS
- 4.4.1 CONTROLLABLE CANONICAL FORM
- 4.4.2 OBSERVABLE CANONICAL FORM
- 4.4.3 TRANSFER FUNCTION (MATRIX) FROM STATE-SPACE EQUATIONS
- 4.5 SOLUTION OF LTI CONTINUOUS-TIME STATE EQUATIONS
- 4.5.1 SOLUTION OF HOMOGENEOUS STATE EQUATION
- 4.5.2 COMPUTING STATE-TRANSITION MATRIX
- 4.5.3 COMPLETE SOLUTION OF STATE EQUATION
- 4.6 STATE-SPACE REPRESENTATION OF DISCRETE-TIME SYSTEMS
- 4.6.1 DEFINITION OF STATE
- 4.6.2 STATE EQUATIONS
- 4.7 STATE-SPACE REPRESENTATION OF DISCRETE-TIME LTI SYSTEMS
- 4.7.1 CONTROLLABLE CANONICAL FORM
- 4.7.2 OBSERVABLE CANONICAL FORM
- 4.8 SOLUTION OF LTI DISCRETE-TIME STATE EQUATIONS
- 4.8.1 SOLUTION OF HOMOGENEOUS STATE EQUATION
- 4.8.2 COMPUTING STATE-TRANSITION MATRIX
- 4.8.3 COMPLETE SOLUTION OF STATE EQUATIONS
- 4.9 CONTROLLABILITY OF LTI SYSTEMS
- 4.9.1 DEFINITION OF CONTROLLABILITY
- 4.9.2 CONTROLLABILITY CONDITION
- 4.10 OBSERVABILITY OF LTI SYSTEMS
- 4.10.1 DEFINITION OF OBSERVABILITY
- 4.10.2 OBSERVABILITY CONDITION
- 5.1 INTRODUCTION
- 5.2 STATE FEEDBACK
- 5.2.1 BASIC CONCEPT
- 5.2.2 POLE-PLACEMENT DESIGN OF SISO SYSTEMS
- 5.3 LQR DESIGN
- 5.3.1 INTRODUCTION
- 5.3.2 SOLUTION OF THE LQR PROBLEM
- 5.3.3 STEADY-STATE ALGEBRAIC RICCATI EQUATION
- 5.4 STATE ESTIMATORS (OBSERVERS) DESIGN
- 5.4.1 INTRODUCTION
- 5.4.2 FULL-ORDER OBSERVER DESIGN
- 5.4.3 REDUCED-ORDER OBSERVER DESIGN
- 5.5 COMBINED STATE ESTIMATION AND CONTROL
- 5.5.1 INTRODUCTION
- 5.5.2 COMBINED CONTROLLER AND OBSERVER
- 6.1 INTRODUCTION
- 6.2 INTERPOLATION OF UNIFORMLY SPACED LOOKUP TABLES
- 6.2.1 LINEAR AND BILINEAR INTERPOLATIONS
- 6.2.2 TRILINEAR INTERPOLATION
- 6.2.3 TETRAHEDRAL INTERPOLATION
- 6.2.4 SEQUENTIAL LINEAR INTERPOLATION
- 6.3 NONUNIFORMLY SPACED LOOKUP TABLES
- 6.3.1 SHEPARD INTERPOLATION
- 6.3.2 MOVING-MATRIX INTERPOLATION
- OF MOVING-MATRIX ALGORITHM
- 6.4 LOOKUP TABLE INVERSE
- 6.4.1 INTRODUCTION
- 6.4.2 INVERSE PRINTER MAP
- 6.4.3 ITERATIVELY CLUSTERED INTERPOLATION
- 6.4.3.1 Selection of Step Size Parameter m
- 6.4.3.2 Algorithm Initialization
- 6.4.4 TETRAHEDRAL TECHNIQUE
- 6.4.5 CONJUGATE GRADIENT APPROACH
- 6.4.6 COMPARISON OF DIFFERENT INVERSION ALGORITHMS
- 6.5 COMPRESSION OF LOOKUP TABLES
- 6.5.1 INTRODUCTION
- 6.5.2 DOWNSAMPLING USING SEQUENTIAL LINEAR INTERPOLATION
- 6.5.3 DYNAMIC OPTIMIZATION ALGORITHM
- 6.5.3.1 One-Dimensional DO Algorithm
- 6.5.3.2 Two-Dimensional DO Algorithm
- 6.5.3.3 Three-Dimensional DO Algorithm
- 6.6 SMOOTHING ALGORITHM FOR MULTIDIMENSIONAL FUNCTIONS
- 6.6.1 INTRODUCTION
- 6.6.2 MULTIDIMENSIONAL SMOOTHING ALGORITHM
- 6.6.2.1 One-Dimensional Smoothing Algorithm
- 6.6.2.2 Two-Dimensional Smoothing Algorithm
- 6.6.2.3 Three-Dimensional Smoothing Algorithm
- 6.6.3 APPLICATION TO PRINTING SYSTEMS
- 7 Three-Dimensional Control of Color Management Systems
- 7.1 INTRODUCTION
- 7.2 IMAGE PATH ARCHITECTURE
- 7.3 PROFILING—A COMPLEX SYSTEM PROBLEM
- 7.3.1 TIGHT COLOR RENDITION REQUIREMENTS
- 7.3.2 GAMUT LIMITATION
- 7.3.3 SMOOTHNESS
- 7.3.5 ENGINE CONDITIONS
- 7.4 CHARACTERIZATION OF COLOR SYSTEMS
- 7.4.1 LEAST-SQUARES ESTIMATION
- 7.4.1.1 A Linear in the Parameters Model
- 7.4.1.2 Recursive Least-Squares Estimation Algorithm
- 7.4.1.3 Piecewise Linear Models
- 7.4.2 PRINCIPAL COMPONENT ANALYSIS-BASED MODEL
- 7.4.2.1 PCA-Based Model in Spectral Space
- 7.4.2.2 PCA-Based Modeling for Adaptive Estimation
- 7.4.2.3 Log-PCA Model (Log-PCA)
- 7.4.2.4 Piecewise Linear PCA Model
- 7.4.2.5 Yule–Nielson Corrected PCA Model
- 7.4.3 NEUGEBAUER MODEL
- 7.4.3.1 Parameterized Model for Neugebauer Weights
- 7.4.3.2 Dot Area Coverages and Neugebauer Weights
- 7.4.3.4 Cellular Neugebauer Model (Lab-NB)
- 7.4.4 DEVICE DRIFT MODEL
- 7.5 GCR SELECTION AND INVERSION
- 7.5.1 A SIMPLE GCR FUNCTION
- 7.5.2 INVERSION OF A THREE-TO-THREE FORWARD MAP
- 7.5.2.1 Inverse by Working on the Printer Model
- 7.5.2.2 Control-Based Inversion
- 7.5.2.3 Inverse by Iterating Directly on the Printer
- 7.5.3 BRIEF REVIEW OF GCR METHODS
- 7.5.4 GCR CONSTRAINED 4-TO-3 INVERSE
- 7.5.4.1 A 4-to-3 Control-Based Inversion
- 7.5.4.2 K-Restricted GCR
- 7.5.4.3 Tricolor GCR
- 7.5.5 GCR RETRIEVAL FROM HISTORICAL PROFILES
- 7.5.6 K-SUPPRESSION METHODS
- 7.6 GAMUT-MAPPING METHODS
- 7.6.1 GAMUT MAPPING WITH RAY-BASED CONTROL MODEL
- 7.6.2 CENTROID CLIPPING
- 7.6.3 SOFT GAMUT MAPPING WITH RAY-BASED CONTROL MODEL
- 7.6.4 GAMUT MAPPING FOR CONSTANT LIGHTNESS AND HUE
- 7.6.5 MERIT-BASED GAMUT MAPPING
- 7.6.6 BLACK POINT COMPENSATION
- 7.7 EVALUATION OF PROFILES
- 7.7.1 GAMUT UTILIZATION AND ROUND TRIP ACCURACY
- 7.7.2 GAMUT CORNER PLOTS AND NEUTRAL RESPONSE
- 7.7.3 VISUAL EVALUATION OF PROFILES
- 7.8 AN EXAMPLE SHOWING HOW TO BUILD MULTIDIMENSIONAL INVERSE LUT
- 8.1 INTRODUCTION
- 8.2 PRINCIPLES OF COLOR MANAGEMENT
- 8.3 ONE-DIMENSIONAL GRAY-BALANCE CALIBRATION
- 8.4 TWO-DIMENSIONAL CALIBRATION
- 8.5.2 GRAY-BALANCED CALIBRATION
- 8.5.3 TWO-DIMENSIONAL CALIBRATION
- 8.6.1 POLE-PLACEMENT DESIGN
- 8.6.2 HIGHLIGHT AND SHADOW CORRECTIONS
- 8.6.2.1 Highlight Corrections
- 8.6.2.2 Shadow Corrections
- 8.6.4 PREDICTIVE GRAY BALANCE
- 8.7 SPOT-COLOR CONTROL
- 8.7.1 GAMUT MAPPING FOR SPOT-COLOR CONTROL
- 8.7.2 GAMUT CLASSES
- 8.7.3 CONTROL ALGORITHM
- 8.7.4 CONTROL ALGORITHM WITH INK LIMITS
- 9.1 INTRODUCTION
- 9.2 PROCESS CONTROL MODELS—A GENERAL CONTROL VIEW
- 9.3 TIME HIERARCHICAL PROCESS CONTROL LOOPS
- 9.4 LEVEL 1 ELECTROSTATIC CONTROL SYSTEM
- 9.4.1 ELECTROSTATIC CONTROLLER DESIGN
- 9.5 STATE SPACE TO TRANSFER FUNCTION CONVERSIONS
- 9.6 LEVEL 2 DEVELOPABILITY CONTROLLER
- 9.6.1 JACOBIAN MATRIX FOR DEVELOPABILITY CONTROL
- 9.7 STEADY-STATE ERROR
- 9.8 DESIGN OF THE GAIN MATRIX
- 9.9 LEVEL 3 CONTROL LOOPS
- 9.9.1 STATIC TRC INVERSION PROCESS
- 9.9.2 CONTROL-BASED TRC INVERSION PROCESS
- 9.10 DEAD BEAT RESPONSE
- 9.11 TC CONTROL LOOP
- 9.11.1 OPEN-LOOP TC MODEL
- 9.11.2 DESIGN OF A TC CONTROL LOOP USING A PI CONTROLLER
- 9.11.4 FEEDFORWARD COMPENSATION FOR IMAGE DISTURBANCE
- 9.12 PROCESS CONTROLS UNDER LIMITED ACTUATION
- 9.13 OPTIMAL CONTROLS FOR SELECTIVE STATES
- 9.14 OPTIMAL MEASUREMENTS
- 10.1 INTRODUCTION
- 10.2 PROCESS MODELS
- 10.2.1 CHARGING MODEL
- 10.2.2 EXPOSURE MODEL
- 10.2.3 DEVELOPMENT MODEL
- 10.2.4 TRANSFER MODEL
- 10.2.5 FUSING MODEL
- 10.2.6 COLOR MODEL
- 10.2.6.1 Sensitivity Analysis of the Model
- 10.3 MODULATION TRANSFER FUNCTIONS
- 10.4 TONE REPRODUCTION CURVE
- 10.5 IMAGE SIMULATION WITH FUSING AND COLOR MODELS
- 10.6 VIRTUAL PRINTER COLOR GAMUT
- 10.7 VIRTUAL PRINTER MODEL TUNING TO AN EXPERIMENTAL PRINTER
- 10.7.1 TUNING TONER MASTER CURVES
- 10.7.2 TUNING OF SINGLE SEPARATION COEFFICIENTS
- 10.7.3 DETERMINATION OF COLOR MIXING COEFFICIENTS {Cji}
- 10.7.4 ONE-DIMENSIONAL CHANNEL-WISE TRC MATCHING
- 10.7.5 TUNING RESULTS
- 10.7.6 SUMMARY
- Appendix A
- Appendix B
- Appendix C
- Index

Analysis and Design

Lalit K. Mestha Sohail A. Dianat

Boca Raton London New York

CRC Press is an imprint of the Taylor & Francis Group, an informa business

CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 33487-2742 © 2009 by Taylor & Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group, an Informa business No claim to original U.S. Government works Printed in the United States of America on acid-free paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number-13: 978-0-8493-3746-8 (Hardcover) This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Library of Congress Cataloging-in-Publication Data Mestha, L. K. Control of color imaging systems : analysis and design / authors, L.K. Mestha and Sohail A. Dianat. p. cm. “A CRC title.” Includes bibliographical references and index. ISBN 978-0-8493-3746-8 (alk. paper) 1. Imaging systems--Automatic control. 2. Color display systems--Automatic control. 3. Digital printing--Automatic control. 4. Image processing--Digital techniques. 5. Color printing. I. Dianat, Sohail A. II. Title. TK8315.M47 2009 681’.62--dc22 Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com and the CRC Press Web site at http://www.crcpress.com 2008054552

Dedication

This book is dedicated to our families, Suhan, Savan, and Veena Mestha, Ahrash Dianat, and Mitra Nikaein

Contents

Preface..................................................................................................................... xv Acknowledgments.................................................................................................. xix

Chapter 1 1.1 1.2

An Overview of Digital Printing Systems........................................... 1

Introduction .................................................................................................... 1 Printing and Publishing System..................................................................... 1 1.2.1 Business Management ........................................................................ 2 1.2.2 Output Production............................................................................... 2 1.2.3 Process Management .......................................................................... 5 1.3 Digital Front End ........................................................................................... 5 1.4 Digital Print Engine (Electrophotographic) ................................................... 6 1.4.1 Image-on-Image and Tandem Print Engines ...................................... 7 1.4.2 Parallel Printing Systems .................................................................... 8 1.5 Evolution of Controls Technology for Digital Printers—Color Controls View.................................................................... 10 1.6 Prepress-Based Processing........................................................................... 13 1.7 DFE-Based Processing................................................................................. 15 1.8 Print Engine-Based Processing .................................................................... 16 References ............................................................................................................... 17 Chapter 2 2.1 2.2 Fundamentals of Digital Image Processing ....................................... 19

2.3 2.4

2.5

Introduction .................................................................................................. 19 Digital Image Formation and Systems ........................................................ 19 2.2.1 Point Spread Function of a Defocused Lens .................................... 20 2.2.2 Point Spread Function of Motion Blur ............................................. 21 2.2.3 Point Spread Function of Human Visual System............................. 22 Optical and Modulation Transfer Functions................................................ 22 Image Sampling and Quantization............................................................... 28 2.4.1 Two-Dimensional Sampling Theorem.............................................. 31 2.4.2 Image Quantization ........................................................................... 34 2.4.2.1 Uniform Quantization ......................................................... 34 2.4.2.2 Signal-to-Quantization Noise Ratio (SQNR)...................... 35 2.4.2.3 Optimum Minimum Mean-Square Error Quantizer............ 36 2.4.2.4 Perceptual Quantization ...................................................... 40 2.4.2.5 Vector Quantization ............................................................ 41 Image Transform.......................................................................................... 46 2.5.1 Two-Dimensional Discrete Fourier Transform................................. 46 2.5.2 Two-Dimensional Discrete Cosine Transform ................................. 57

vii

viii

Contents

Two-Dimensional Hadamard Transform ...................................... 58 2.5.3.1 Inverse Hadamard Transform........................................ 59 2.6 Image Filtering........................................................................................... 60 2.6.1 Design of 2-D FIR Filters............................................................. 64 2.7 Image Resizing........................................................................................... 68 2.7.1 Deﬁnition of Sampling Rate Conversion ..................................... 68 2.7.2 Upsampling by Factor of P .......................................................... 69 2.7.3 Downsampling by Factor of Q ..................................................... 70 P 2.7.4 Sampling Rate Conversion by a Factor of Q ................................ 72 2.7.5 Examples of Low-Pass Filters Used for Sampling Rate Conversion............................................................................ 72 2.8 Image Enhancement................................................................................... 77 2.8.1 Unsharp Masking.......................................................................... 77 2.8.2 Image Histogram........................................................................... 78 2.8.3 Histogram Equalization................................................................. 80 2.9 Image Restoration ...................................................................................... 82 2.9.1 Wiener Filter Restoration.............................................................. 83 2.10 Image Halftoning ....................................................................................... 85 2.10.1 Error Diffusion Algorithm ............................................................ 88 Problems.................................................................................................................. 91 References ............................................................................................................... 97 Chapter 3 3.1 3.2 Mathematical Foundations................................................................. 99

2.5.3

3.3 3.4 3.5

3.6

3.7 3.8

3.9

Introduction ................................................................................................ 99 General Continuous-Time System Description ......................................... 99 3.2.1 Solution of Constant-Coefﬁcients Linear Differential Equations ................................................................. 100 Laplace Transform ................................................................................... 102 3.3.1 Inverse Laplace Transform ......................................................... 104 General Linear Discrete-Time Systems ................................................... 107 3.4.1 Solution of Constant-Coefﬁcients Difference Equations............ 108 z-Transform .............................................................................................. 110 3.5.1 Properties of z-Transform ........................................................... 113 3.5.2 Inverse z-Transform .................................................................... 120 3.5.3 Relation between the z-Transform and the Laplace Transform ... 125 Discrete-Time Fourier Transform ............................................................ 127 3.6.1 Properties of Discrete-Time Fourier Transform ......................... 127 3.6.2 Inverse DTFT.............................................................................. 128 Two-Dimensional z-Transform ................................................................ 129 Two-Dimensional Discrete-Space Fourier Transform ............................. 131 3.8.1 Properties of 2-D DSFT.............................................................. 132 3.8.2 Inverse 2-D DSFT....................................................................... 132 Eigenvalues and Eigenvectors ................................................................. 134 3.9.1 Deﬁnition of Eigenvalue and Eigenvector.................................. 134 3.9.2 Product and Sum of Eigenvalues................................................ 137

Contents

ix

3.9.3 Finding Characteristic Polynomial of a Matrix .......................... 138 3.9.4 Modal Matrix .............................................................................. 139 3.9.5 Matrix Diagonalization ............................................................... 140 3.9.6 Deﬁnite Matrices......................................................................... 142 3.10 Singular Value Decomposition ................................................................ 144 3.10.1 Matrix Norm ............................................................................... 146 3.10.2 Principal Components Analysis.................................................. 150 3.11 Matrix Polynomials and Functions of Square Matrices .......................... 155 3.11.1 Matrix Polynomial ...................................................................... 156 3.11.2 Inﬁnite Series of Matrices........................................................... 156 3.11.3 Cayley–Hamilton Theorem......................................................... 157 3.11.4 Function of Matrices................................................................... 159 3.11.4.1 Cayley–Hamilton Technique..................................... 159 3.11.4.2 Modal-Matrix Technique .......................................... 161 3.11.5 Matrix Exponential Function eAt ................................................ 163 3.11.6 Computing eAt Using Laplace Transform ................................. 165 3.11.7 Matrix Exponential Function Ak ................................................. 166 3.12 Fundamentals of Matrix Calculus............................................................ 168 3.12.1 Derivatives of a Scalar Function with Respect to a Vector ....... 168 3.12.2 Derivatives of Quadratic Functions ............................................ 170 3.12.3 Derivative of a Vector Function with Respect to a Vector ........ 172 Problems................................................................................................................ 172 References ............................................................................................................. 176 Chapter 4 4.1 4.2 4.3 State-Variable Representation.......................................................... 177

4.4

4.5

4.6

Introduction .............................................................................................. 177 Concept of States ..................................................................................... 177 State-Space Representation of Continuous-Time Systems...................... 177 4.3.1 Deﬁnition of State....................................................................... 177 4.3.2 State Equations of Continuous-Time Systems ........................... 178 4.3.3 State-Space Equations of Electrical Systems.............................. 179 4.3.4 State-Space Equations of Mechanical Systems .......................... 182 State-Space Representation of General Continuous LTI Systems .......... 185 4.4.1 Controllable Canonical Form...................................................... 186 4.4.2 Observable Canonical Form ....................................................... 186 4.4.3 Transfer Function (Matrix) from State-Space Equations............ 187 Solution of LTI Continuous-Time State Equations ................................. 188 4.5.1 Solution of Homogeneous State Equation.................................. 188 4.5.2 Computing State-Transition Matrix ............................................ 189 4.5.3 Complete Solution of State Equation ......................................... 191 State-Space Representation of Discrete-Time Systems ........................... 193 4.6.1 Deﬁnition of State....................................................................... 193 4.6.2 State Equations............................................................................ 194

x

Contents

4.7

State-Space Representation of Discrete-Time LTI Systems.................... 195 4.7.1 Controllable Canonical Form...................................................... 195 4.7.2 Observable Canonical Form ....................................................... 196 4.7.3 Transfer Function (Matrix) from State-Space Equations..................................................................................... 197 4.8 Solution of LTI Discrete-Time State Equations ...................................... 198 4.8.1 Solution of Homogeneous State Equation.................................. 198 4.8.2 Computing State-Transition Matrix ............................................ 199 4.8.3 Complete Solution of State Equations........................................ 201 4.9 Controllability of LTI Systems ................................................................ 203 4.9.1 Deﬁnition of Controllability ....................................................... 203 4.9.2 Controllability Condition ............................................................ 204 4.10 Observability of LTI Systems.................................................................. 205 4.10.1 Deﬁnition of Observability ......................................................... 206 4.10.2 Observability Condition.............................................................. 206 Problems................................................................................................................ 209 References ............................................................................................................. 213 Chapter 5 5.1 5.2 Closed-Loop System Analysis and Design ..................................... 215

Introduction .............................................................................................. 215 State Feedback ......................................................................................... 215 5.2.1 Basic Concept ............................................................................. 215 5.2.2 Pole-Placement Design of SISO Systems................................... 219 5.2.3 Pole-Placement Design of Multiple-Input Multiple-Output (MIMO) Systems ............................................. 224 5.2.4 Relationship between Poles and the Closed-Loop System Response ........................................................................ 227 5.3 LQR Design ............................................................................................. 227 5.3.1 Introduction ................................................................................. 227 5.3.2 Solution of the LQR Problem..................................................... 229 5.3.3 Steady-State Algebraic Riccati Equation.................................... 233 5.4 State Estimators (Observers) Design ....................................................... 234 5.4.1 Introduction ................................................................................. 234 5.4.2 Full-Order Observer Design ....................................................... 234 5.4.3 Reduced-Order Observer Design ................................................ 238 5.5 Combined State Estimation and Control ................................................. 240 5.5.1 Introduction ................................................................................. 240 5.5.2 Combined Controller and Observer............................................ 240 Problems................................................................................................................ 244 References ............................................................................................................. 247 Chapter 6 6.1 6.2 Interpolation of Multidimensional Functions .................................. 249

Introduction .............................................................................................. 249 Interpolation of Uniformly Spaced Lookup Tables................................. 250 6.2.1 Linear and Bilinear Interpolations .............................................. 250 6.2.2 Trilinear Interpolation ................................................................. 254

Contents

xi

6.2.3 Tetrahedral Interpolation................................................................. 257 6.2.4 Sequential Linear Interpolation....................................................... 262 6.3 Nonuniformly Spaced Lookup Tables....................................................... 264 6.3.1 Shepard Interpolation...................................................................... 264 6.3.2 Moving-Matrix Interpolation .......................................................... 267 6.3.3 Recursive Least-Square Implementation of Moving-Matrix Algorithm.......................................................... 269 6.4 Lookup Table Inverse ................................................................................ 270 6.4.1 Introduction..................................................................................... 270 6.4.2 Inverse Printer MAP ....................................................................... 270 6.4.3 Iteratively Clustered Interpolation .................................................. 272 6.4.3.1 Selection of Step Size Parameter m .................................. 273 6.4.3.2 Algorithm Initialization ..................................................... 274 6.4.4 Tetrahedral Technique .................................................................... 274 6.4.5 Conjugate Gradient Approach ........................................................ 275 6.4.6 Comparison of Different Inversion Algorithms.............................. 276 6.5 Compression of Lookup Tables................................................................. 277 6.5.1 Introduction..................................................................................... 277 6.5.2 Downsampling Using Sequential Linear Interpolation .................. 278 6.5.3 Dynamic Optimization Algorithm .................................................. 278 6.5.3.1 One-Dimensional DO Algorithm ...................................... 278 6.5.3.2 Two-Dimensional DO Algorithm ..................................... 280 6.5.3.3 Three-Dimensional DO Algorithm ................................... 281 6.6 Smoothing Algorithm for Multidimensional Functions ............................ 286 6.6.1 Introduction..................................................................................... 286 6.6.2 Multidimensional Smoothing Algorithm ........................................ 288 6.6.2.1 One-Dimensional Smoothing Algorithm .......................... 288 6.6.2.2 Two-Dimensional Smoothing Algorithm.......................... 289 6.6.2.3 Three-Dimensional Smoothing Algorithm........................ 290 6.6.3 Application to Printing Systems ..................................................... 293 Problems................................................................................................................ 295 References ............................................................................................................. 301

Chapter 7 7.1 7.2 7.3

Three-Dimensional Control of Color Management Systems .......... 303

7.4

Introduction ................................................................................................ 303 Image Path Architecture ............................................................................ 303 Proﬁling—A Complex System Problem.................................................... 305 7.3.1 Tight Color Rendition Requirements.............................................. 305 7.3.2 Gamut Limitation............................................................................ 306 7.3.3 Smoothness ..................................................................................... 306 7.3.4 ICC Workﬂow................................................................................. 307 7.3.5 Engine Conditions........................................................................... 307 Characterization of Color Systems ............................................................ 308 7.4.1 Least-Squares Estimation................................................................ 308 7.4.1.1 A Linear in the Parameters Model.................................... 309

xii

Contents

7.4.1.2 Recursive Least-Squares Estimation Algorithm ............... 310 7.4.1.3 Piecewise Linear Models .................................................. 313 7.4.2 Principal Component Analysis-Based Model ................................ 321 7.4.2.1 PCA-Based Model in Spectral Space ............................... 321 7.4.2.2 PCA-Based Modeling for Adaptive Estimation ............... 327 7.4.2.3 Log-PCA Model (Log-PCA) ............................................ 329 7.4.2.4 Piecewise Linear PCA Model ........................................... 329 7.4.2.5 Yule–Nielson Corrected PCA Model ............................... 330 7.4.3 Neugebauer Model.......................................................................... 331 7.4.3.1 Parameterized Model for Neugebauer Weights ................ 332 7.4.3.2 Dot Area Coverages and Neugebauer Weights ................ 335 7.4.3.3 Estimation of Dot Area Coverages Using Least Squares.......................................................... 337 7.4.3.4 Cellular Neugebauer Model (Lab-NB) ............................. 339 7.4.4 Device Drift Model......................................................................... 344 7.4.4.1 Autoregressive (AR) Model Applied to Printer Drift Prediction.................................................................. 344 7.4.4.2 Vector Autoregressive Model Applied to Printer Drift Prediction.................................................................. 347 7.5 GCR Selection and Inversion .................................................................... 350 7.5.1 A Simple GCR Function ................................................................ 351 7.5.2 Inversion of a Three-to-Three Forward Map.................................. 353 7.5.2.1 Inverse by Working on the Printer Model........................ 354 7.5.2.2 Control-Based Inversion ................................................... 355 7.5.2.3 Inverse by Iterating Directly on the Printer ...................... 359 7.5.3 Brief Review of GCR Methods ...................................................... 362 7.5.4 GCR Constrained 4-to-3 Inverse .................................................... 364 7.5.4.1 A 4-to-3 Control-Based Inversion .................................... 365 7.5.4.2 K-Restricted GCR ............................................................. 366 7.5.4.3 Tricolor GCR .................................................................... 377 7.5.5 GCR Retrieval from Historical Proﬁles.......................................... 379 7.5.6 K-Suppression Methods .................................................................. 382 7.6 Gamut-Mapping Methods .......................................................................... 384 7.6.1 Gamut Mapping with Ray-Based Control Model .......................... 385 7.6.2 Centroid Clipping ........................................................................... 392 7.6.3 Soft Gamut Mapping with Ray-Based Control Model................... 393 7.6.4 Gamut Mapping for Constant Lightness and Hue.......................... 395 7.6.5 Merit-Based Gamut Mapping ......................................................... 396 7.6.6 Black Point Compensation ............................................................. 398 7.7 Evaluation of Proﬁles................................................................................. 399 7.7.1 Gamut Utilization and Round Trip Accuracy ................................ 399 7.7.2 Gamut Corner Plots and Neutral Response .................................... 400 7.7.3 Visual Evaluation of Proﬁles .......................................................... 406 7.8 An Example Showing How to Build Multidimensional Inverse LUT...... 412 Problems................................................................................................................ 421 References ............................................................................................................. 422

......................................................................................2............................................... 513 ....................... 442 8.. 459 8.......... 450 8................................. 436 8...............3 Control Algorithm..............................4 Internal Process Controls. 467 8..........3 9................. 453 8...2 9..............1 Pole-Placement Design .........................5 State Space to Transfer Function Conversions...............................6..7 Spot-Color Control......9.....6........7. 494 9...........................8 Design of the Gain Matrix ............... 505 9..................Contents xiii Chapter 8 8........................................................................ 433 Two-Dimensional Calibration....4 8...5............... Two-Dimensional...................................... 435 One-Dimensional and Two-Dimensional Printer Calibration Using Printer Models ................... and Spot-Color Management and Control Methods ............................................................................................................... 491 9.............6.......1 Highlight Corrections.......................................................................1 Jacobian Matrix for Developability Control.............................. 472 Time Hierarchical Process Control Loops.................. 436 8.................................................................................... 483 9...2 Gamut Classes.................................. 469 References ..............7................1 8...................9 Level 3 Control Loops...............2 Control-Based TRC Inversion Process .....1 Static TRC Inversion Process ..... 438 8....................................................................6.............................................................................................................10 Dead Beat Response ....... 454 8.............................................................................................................2 Highlight and Shadow Corrections............5.......... 477 9........................2 Gray-Balanced Calibration...6 One-Dimensional and Two-Dimensional Printer Calibration with State-Feedback Methods.................................................. 501 9..................1 Electrostatic Controller Design ................................................................................................................2 8.................................3 Two-Dimensional Printer Calibration with State-Feedback Methods....6 Level 2 Developability Controller ............. 455 8...........7 Steady-State Error ........................... 464 8...6........ 488 9.............................................................................................................................................4 Control Algorithm with Ink Limits ........ 471 Process Control Models—A General Control View ........... 431 Principles of Color Management ................1 9...............6......................................................9...... 449 8..............3 Two-Dimensional Calibration.............1 One-Dimensional Channel-Wise (Independent) Calibration......1 Gamut Mapping for Spot-Color Control ..........................4 Predictive Gray Balance .5 One-Dimensional......................................... 486 9.. 432 One-Dimensional Gray-Balance Calibration ...................................2.............. 471 Introduction .............................................................................................. 445 8............................ 464 8.................2 Shadow Corrections ........... 457 8....................3 8........................5................................ 477 Level 1 Electrostatic Control System ........................................ 470 Chapter 9 9...................................................... 509 9..............7....... 467 Problems............... 431 Introduction ....................7......................................................................... 496 9..........4.........................6.............

...........................................2.................................7............. 612 10.... 613 10.11........3 Design of a TC Control Loop with a Time Delay Using a PI Controller ............................................ 618 Problems.2 Tuning of Single Separation Coefﬁcients..............1 Charging Model ..................4 Transfer Model............................4 Tone Reproduction Curve........................................... 515 9.... 562 10................................................................ 549 References ...........................5 Fusing Model ..............................................................2.......................... 597 10........................... 530 9..............................................................................11....................................................... 641 Appendix C ........14 Optimal Measurements ................... 558 10................................ 557 Introduction . 645 Index................................................................................3 Development Model.........................................................................2 Printing System Models ........... 615 10.....................................................7 Virtual Printer Model Tuning to an Experimental Printer....................................................................................................................7....11........................................... 608 10.......7...... 569 10........ 605 10..................................................................................2 Exposure Model ..........................7...............................................................................................................5 Tuning Results ......................... 593 10.............3 Modulation Transfer Functions.......................................................4 Feedforward Compensation for Image Disturbance ................................................................................................... 584 10............................................... 543 Problems.6 Virtual Printer Color Gamut .......................................................................7.. 647 ........................ 517 9................ 515 9.......... 606 10..................................................11 TC Control Loop ........7..................................2 Design of a TC Control Loop Using a PI Controller.............................................................................11......13 Optimal Controls for Selective States......................................4 One-Dimensional Channel-Wise TRC Matching .......................................... 525 9.................................................................... 623 Appendix B .........................6............... 577 10..........2.............1 Tuning Toner Master Curves........2..............................................................................................1 10................................2......................2............... 552 Chapter 10 10............................. 526 9........ 618 References ......1 Open-Loop TC Model ..1 Sensitivity Analysis of the Model.....xiv Contents 9.................................... 619 Appendix A .....................6 Color Model.5 Image Simulation with Fusing and Color Models . 617 10.............................. 610 10...6 Summary .....2.............. 522 9...........11...........5 Design of TC Control Loop with State Feedback Controller and State Estimator....................... 557 10................................................................................ 586 10. 557 Process Models ......................................................... 610 10...................3 Determination of Color Mixing Coefﬁcients {Cji} .............................................................................................................12 Process Controls Under Limited Actuation .......... 539 9..........

toner mass. At a system level. CMYK primaries etc. stocks and environment. to some extent. and closed loop feedback controls. The quality and productivity issues of these devices are addressed using a variety of new technologies including optical sensing. In this process. no design books have been written on this subject matter. it is also xv . or a student. color management and control. at a senior level or at a ﬁrst-year graduate level. impossible. Also. and the principles associated with electronic imaging. it can also be used as a textbook for an introductory course in printing. digital control. The electrostatic image is then developed with a thermoplastic powder containing charged pigment that is transferred and fused to paper under heat and pressure. Someone new to the ﬁeld. the image content controls the amount of light that selectively discharges a uniformly charged photoreceptor material with a laser or light emitting diodes to form an image. Unlike off-set printers. print quality defects and subtle variations in output are more noticeable in color printing than in monochrome printing. theoretical knowledge of the physical printing process. device technology. color printing etc.. digital imaging systems. the ability to produce accurate and pleasing color across numerous output devices is extremely complex and. Practical feedback control systems used in digital production printers touch on a range of interconnected subsystems. laser intensity. should have a good foundation in the major disciplines used for managing & controlling color in digital production printers.) for developing real-time. In our opinion. So. This book brings together the numerous complex disciplines associated with digital color printing and presents a technical story with proper mathematical rigor and design examples.. a practitioner. Although there are numerous digital printers serving today’s market. color is rendered on different imaging devices (variety of printing and display devices) with varying color capabilities.g. imaging. to the best of our knowledge. We have attempted to keep the mathematics at a moderate level. digital printers offer many new actuators that give control engineers access to numerous process steps (e. Since the process tends to vary more over time. who could be a researcher. which is used in the process of laser printing. This is further made difﬁcult due to variation in the workﬂow requirements. Although an extensive number of patents and conference papers have been written on the control of digital color printing.Preface Digital color printing technology offers many new avenues and opportunities for rendering color pages on demand at a lower run cost as compared with conventional printing. Many digital printers are based on electrophotographic technology. The objective of this book is to provide a good understanding of fundamental techniques and push the frontier of this ﬁeld. As a result. many new challenges must be overcome to improve the output quality and enable further growth and opportunity. closed-loop algorithms and architectures that can be self-tuned using various sensors along the print path and self-corrected for process variations and uncertainties. a good basic understanding of the end-to-end color printing process may be helpful and at a subsystem level.

optical and modulation transfer function. the solution of state equations. Design techniques such as pole placement for single-input single-output (SISO) as well as multi-input multi-output systems (MIMO) and linear quadratic regulator design are among the topics that are covered in this chapter. and the pulse transfer function are described. closed-loop controller design for charging and development systems are presented. workﬂows of data. It is assumed that the readers have a basic understanding of color. the basics of 1-D and 2-D tone control of color management systems are discussed. sequential linear interpolation. image transform. Shepard. The contents of the book are outlined as follows: Chapter 1 provides an overview of digital printing systems. and stability of linear systems are major topics discussed in this chapter. Chapter 7 covers the 3-D control of color management systems with International Color Consortium (ICC) proﬁles generated using gray-component replacement (GCR) constraints. and spot color control methods are presented. State variable representation of imaging systems. digital front-end system. dynamic optimization and 1-D. Topics such as differential and difference equations (DE). the numerical solution of DE. various empirical and ﬁrst principle based Neugebauer models are covered in detail. the z-transform and its properties. State variable representation. sampling and quantization. where precision color controls and sensing are needed to achieve offset quality color images. the relationship between the z-transform and the Laplace transform. In this chapter. They use algorithms and capabilities far beyond standard printing and copying machines. The material covered in this book is based on our 27 years of combined experience in applying control algorithms to modern digital printing systems. this book can serve as a reference for instructors or a self-learning book for practicing engineers. tetrahedral interpolation. Techniques such as trilinear interpolation. including topics like convolution and Fourier analysis. and applied linear algebra. control-based 1-D and 2-D tone management. This book also captures some of our combined experience of over 30 years in teaching fundamental courses in engineering. Chapter 2 covers some fundamental topics. In Chapter 9. a detailed modeling and analysis of internal closed-loop process controls. Additionally. 3-D smoothing algorithms are described. linear system theory (both continuous and discrete). especially in the areas of digital signal=image processing. controllability. which have appeared as products for automating color consistency in digital presses. and the major elements in an end-to-end production print path. Chapter 4 deals with state variable techniques used to analyze continuous and discrete linear systems. Chapter 3 covers the mathematical tools needed for the analysis and design of feedback control systems. Current best practices used in toner concentration . moving matrix. and image de-noising. control-based inversion and control-based gamut mapping approaches. 2-D. Chapter 6 explains different techniques used for interpolation of multidimensional functions. iterative clustered interpolation (ICI) algorithm. In Chapter 8. color. image coding. modern control theory. which form the cornerstone of digital image processing. observability. state transition matrix. Techniques for closed-loop linear system analysis and design are covered in Chapter 5. such as digital image formation. This book largely discusses the imaging systems developed over recent years.xvi Preface suitable for a graduate level course in imaging and computer applications.

and fusing subsystems are modeled in Chapter 10. including measurement of spectral functions of color are considered beyond the scope of this book. While constructing end-to-end system models with process and subsystem parameters.Preface xvii system modeling with time delay & controller design for time delay systems are covered. state feedback and linear quadratic regulator design so that the mathematics is reasonably simple and can be taught at a senior undergraduate or a ﬁrst year graduate level class in color. Mestha Sohail A. The control theory and methods presented in this book are state-of-the art for color printing systems. Our approach shown in Chapter 10.g.. Dianat . Formulations and illustrations presented emphasize simplicity so that the readers can easily understand the concepts and use them in their systems in the same way we did for high-end printers. although elementary. Lalit K. algorithms and methods in imaging systems to achieve full autonomy. Components of a digital laser printing system. accurate. and offset-like output at lower cost through automation. faculty and practicing engineers to explore more advanced approaches for designing future imaging systems that compare and contrast with experimental data when they become available. transfer. We hope this book will bridge the gap between current and future theoreticians and practitioners. We deliberately limit the control theory to MIMO pole placement. Tuning of model parameters is a rich area for applying modern system identiﬁcation methods. as well as generate new ideas. Emphasis is on essential theoretical design principles and algorithms needed to build high quality. is considered useful to bring some reality to simulations since most control systems designed today are ﬁrst simulated before a real test is done. control or computer engineering disciplines. System models are presented in Chapter 10 to provide a nonlinear simulation platform for students. charging. we tried to capture the meaningful & essential behavior of subsystems in terms of parameters accessible for designing control systems. imaging. development. toner density measurements on photoreceptors). exposure. Optical sensing methods used for measuring various process outputs (e.

.

my research would have been different and the book would have been written on a different topic. Professor Richard Talman of Cornell University. Several Industry-University collaborations were initiated since I started writing the book in 1996: a National Science Foundation’s Grant Opportunity for Academic LIason (GOALI). MIMO poleplacement design. Mihalyov. Applying all of these methods and algorithms is truly a great success for me. he exposed me to the complex world of digital color printing. before I go any further. and Professor Kai Yeung of the University of Texas at Arlington for injecting the seeds of physics and controls into my mind before I moved on to Xerox Research and realized the need for sophisticated learning when designing large scale physical systems. optimal linear estimators. Industry-University interactions and teaching as an adjunct professor at the University of Texas at Arlington and the Rochester Institute of Technology (RIT) greatly improved my ability to internalize and leverage existing methods and algorithms for real world system applications. input-output experimental processes. Thanks to the support of his staff. as well as Peter A. Michael R. Bob Webber of Fermi National Accelerator Laboratory. one of 15 NYSTAR sponsored Centers for Advanced Technology (CATs) in the area of control theory and electronic imaging. a senior research fellow. Charles B. Kenneth J.Acknowledgments I would like to thank Graham Rees of Rutherford Appleton Laboratory. state variable methods. Therefore. and I would like to acknowledge the support of all the faculty members and students who worked with me for all of those years. linear MIMO state feedback. Furst. I would like to recognize my managers. who always gave me the support I needed for carrying out the research and producing the text book. who have supported me at various stages in my research career while developing my knowledge in color and xerographic systems. Thieret. dynamic optimization. Duke. It was Dr. anti-windup compensators for saturation. I would like to acknowledge and express special gratitude to my Senior Management. Crean. design approaches for time delay systems. Debbie Wickham. and Lisa Purvis. Multi-dimensional smoothing. New York State’s Center for Electronic Imaging Systems (CEIS). If it wasn’t for him. Norm W. Zeck. I would like to especially xix . Sophie Vandebroek. from approval to completion. Tracy E. Charles B. These academic relationships led to many important research results. Steve Hoover and Steve Bolte. I would like to express my most sincere gratitude and appreciation to Dr. and system identiﬁcation are just a few of the methods that are now routinely used in our research regarding next generation systems. while trying to meet their business objectives. principal component analysis. Hannaway. and working with product development groups so that our research led to the creation of value for our customers. Duke who brought me to Xerox Corporation while I was looking for a new job in Texas and pondering my next big challenging control application. William J. singular value decomposition.

our student intern at RIT. whose experience and intuition helped us to understand color. Special thanks goes to Prudhvi Krishna Gurram. Without his focused effort and numerous suggestions. technicians. and Mitra Nikaein. but also contributed to the tuning of printing system models and read and edited the entire manuscript. a Principal Scientist at Xerox Corporation. many errors may have gone to press unnoticed. Dianat . and developers at Xerox Corporation who contributed to the growth of the technology with us over the years and made it useful to our customers. Nikolaos M. for their support and understanding when dedicating our personal time on this project. Lastly. who carefully edited the proof line by line. Mestha We would like to thank CEIS and Xerox Corporation for supporting our research in the areas of control for imaging and printing applications over the past ten years. A special thanks to all the administrative staff. Lalit K. We thank Barbara Zimmerli. Freris and Kunal Srivatsava of University of Illinois at Urbana-Champaign also worked in the early stages with us while developing the Printing System models. Writing a book of this magnitude is a major undertaking and producing the material for writing it is even harder. scientists. without his efforts and guidance. our colleague at Xerox Research. color transforms and ICC proﬁles particularly well. Maltz. Savan. who greatly helped with the administrative support while completing the ﬁnal manuscript. Mestha Sohail A. Very special thanks to Palghat Ramesh. Finally. a PhD student at RIT. who helped us in developing a tensor based multidimensional smoothing algorithm that is covered in this book. We would also like to thank Bruce Brewington. it is Jack G. Ahrash Dianat. Kaufman of the Xerox Special Information Systems group for supporting the early injection of our recent research solutions into products. we would not have created the material for Chapter 10. Lalit K. a Principal Scientist at Xerox Research. Elliot. Fisher and Paul A. Special thanks goes to Martin S.xx Acknowledgments thank Peter S. Suhan. and Veena Mestha. we would like to thank our families. who not only got his MSEE degree while working on the CAT funded research project. engineers.

and versioning. valued for newspapers. personalization. and screen-printing use plates or some other form of image carrier. and (3) the beneﬁts of digital data brought additional value.1 An Overview of Digital Printing Systems 1. often called ‘‘offset printing. gravure. ﬂexography.’’ is the dominant printing process in the industry. and print registration systems. we describe some of the key elements of an end-to-end digital printing and publishing system used for the production of high print volumes and the management of complex print jobs.2 PRINTING AND PUBLISHING SYSTEM Printing and publishing is a large industry composed of many shops. 1. that vary in size.1 INTRODUCTION The printing industry includes a number of segments [1] including commercial printing and publishing. In this chapter. These shops use equipments based on a variety of printing methods. edited. that is difﬁcult to create 1 . and consistent print quality are used in packaging and printing of periodicals. directories. and assembled into laid-out pages forming a complete job. This overview will help the reader to better understand the functional and processing system-level components involved when designing optimal printing systems. graphics. Flexography produces vibrant colors with little rub-off qualities. Unlike digital printers. the traditional offset press does not allow the changing of pixels on page boundaries while printing. and controls technology has enabled advanced digital color printing and publishing systems for the ofﬁce and production. (2) the printing technology can handle digital stream of data. Lithography. ﬂexible pagination and formats. letterpress. Lithography. This led to variable data printing with a lower cost short run. imaging. for example. The integration of computing. and often robots are used to move parts in and out of the presses in print shops [2]. ‘print-on-demand’. and books. This limitation brings new challenges and opportunities to the digital printing and publishing value chain. Some publishing processes became digital for several reasons: (1) variable information electronic documents containing fragments of text. and digital or electronic printing such as electrostatic or ink-jet is plateless. In offset printing presses. the press control system controls and monitors the ink. water. and images from either the electronic or the scanned input stream can be merged. Gravure’s highquality reproduction.

Aspects related to handling the shipment of the ﬁnished job. In the creative stage. multiple workstations are networked together to serve as the publishing desktop for generating. the prepress stage (see Figure 1. For low-volume. and assigning job-tracking numbers for monitoring purposes. manipulating. print (press). Understanding some of the key steps involved in the printing and publishing workﬂow can shed some light on the complexity of the system. Electronic documents are then sent to the prepress area for further processing and assembly. 1. Since documents could be of various formats and color spaces (RGB=CMYK).1 BUSINESS MANAGEMENT Business management tasks (the top portion of Figure 1. For high-volume.2 Control of Color Imaging Systems: Analysis and Design with offset printing. and (3) process management=supporting functions (see Figure 1. prepress. demographic data. etc. The typical workﬂows used in both areas of printing (digital and offset) can be divided into three main components: (1) business management. Image capture from scanner or other document input devices are done in this stage. and billing are an integral part of business management functions. paper handouts. Pricing is estimated for the order and compared with the actual cost of running the job.2 OUTPUT PRODUCTION After the order taking step.2. editing. Workﬂows (various steps required all the way from receiving the orders in a print shop to the production of a job in ﬁnished form) are generally unique to each print shop. It includes not just the actual production steps.2) encompasses all the steps involved in creating a digital electronic master. and integrating multimedia content. Once captured. concepts and drawings are developed. illustrations. documents are designed by assembling the content using various layout tools.1 for a block diagram view of the workﬂow process). (2) output production. and variable information documents. Similarly. Generally speaking. the customer data is processed in the creative stage before sending it to prepress. archival.1) involve taking orders from customers. the images—stored electronically in . multishipment. 1. and fulﬁll stages. text. managing. digital still and=or video cameras offer the user the ability to capture single snapshots and=or video imagery. Order information may include artwork. variable information rules. Scanners are usually connected to the workstations to convert hardcopy documents (photos on ﬁlm.2.1). proper design choices are required before converting the documents to the language required by the raster image processor (RIP). archiving. design=layout. etc.) to electronic form by utilizing a variety of image scanning software packages. A project manager handles the job tracking and billing issues. ﬁnish. In a typical prepress system. Customers and sales=service representatives are involved at the order-taking stage. sheet-oriented documents. which converts a document’s strings of character codes to pixels. Customer data is directly sent to the prepress area. This category can be further divided into creative. single-shipment. customer data ﬂow into the output production stage (the middle portion of Figure 1. etc. but all the necessary supporting tasks like billing. offset presses are used.

1 An illustration of the typical print shop workﬂow. 3 .Customer Project manager Job/order tracking Close order Bill Customer data Sales/service rep Take order Business management Proof Press operator Finisher Prepress Create plates Press Mailer driver Approve proof Customer data Prepress An Overview of Digital Printing Systems Creative Creative Prepress Prepress Finish Fulfill Output production Job components Order Process step Order notify target process Schedule Control status Archive & retrieve Inputs & outputs Manage shop/order/job Project manager Process management/supporting functions FIGURE 1.

). Freehand. etc. are usually generated by a scanner or digital camera (still or video). they are assembled into jobs. Trapping is done to the documents to compensate for the small amount of misregistration in the printing system. a raster format (JPEG. checking for fonts. Once the documents have been designed on the desktop. The prooﬁng is sometimes done on another digital printer and often on a wellcalibrated monitor. line art. unnecessary white gaps may appear between two colors that are supposed to be touching. The assembled job is then sent for prooﬁng without actually printing it on the press. inserting=editing multiple pages are done in the prepress area. Without trapping.2 High level steps in the prepress area and DFE for digital printing. Quark XPress. the job is further checked to handle corrections in the prepress area. etc. arrows. setting the right screens (halftone dots).. TIF.4 Raster data Scanner Digital camera … Control of Color Imaging Systems: Analysis and Design Workstation Image processing Color management Document generation Printer drive Digital front end (DFE) Postscript C Image processing Enhancement Noise removal Segmentation Color management Calibration Color rendition dictionary (CRD) M Y K FIGURE 1. and so on. Raster data. boxes. on the other hand. as created by applications like Microsoft Word.)—can be integrated into documents using a variety of desktop publishing software packages such as Microsoft Ofﬁce (Word. If problem pages are found within the prooﬁng stage. Vector data includes line drawings. . etc. PowerPoint. Supported stocks are entered to meet the customer order. Most documents contain a mixture of vector and raster type data. Deﬁning spot=process colors.

For postscript images. For digital printing. format. rasterization. and previewed=proofed on a monitor or on a workgroup digital printer. bundling. During the RIP. where processing by the user may be independent of the print engine. For OOR to be effective. color proﬁles (e. the jobs are submitted to the press—in page description language (PDL) format such as postscript (PS) or PDF—through the digital front end (DFE) for further processing and printing.3 PROCESS MANAGEMENT Process management (the lower portion of Figure 1. PS interpreter) to identify the commands found in the PDL. color management.g. segmentation algorithms must be utilized to identify the objects of interest.. 1. and quality the job is submitted to the print queue. color. In the ﬁnishing (post-press) stage. magenta (M).2) and (b) a print engine (see Figure 1. the printed documents take on their ﬁnal form through cutting.. collating.1) tasks include scheduling production. Job notiﬁcation in the event a new job arrives for production. and organizing ﬁle folders and servers related to all jobs components and job archivals. For high-volume offset jobs.3). Some DFEs employ object-oriented rendering (OOR) algorithms intended to enhance the color reproduction by utilizing custom proﬁles for speciﬁc image objects such as a ‘‘skin’’ proﬁle for ﬂeshtone or a ‘‘sky’’ proﬁle for a blue sky background. which may be a book. industry standard source proﬁles are used to transform RGB images to a device-independent form like L*a*b* or standard Web offset printing (SWOP) CMYK ﬁles to device-independent form. booklets nicely cut and bound. and customer communication handling is another key task executed in this step. 1. 600 dpi). to CMYK color separations to be printed by the engine. job intervention. it is a . folded=collated sheets.An Overview of Digital Printing Systems 5 evaluated for color quality.3 DIGITAL FRONT END The printing stage normally contains: (a) a DFE (see Figure 1. and=or packaging operations prior to shipment to customers. a DFE or a network of DFEs from multiple vendors are used to convert the electronic ‘‘master’’ documents or job (through a series of image processing applications such as trapping. yellow (Y). the transformation may be direct between device-speciﬁc forms to printer-speciﬁc form. The ﬁnishing work has a major impact on the ﬁnal product’s quality. To this effect. International Color Consortium [ICC]) comprising of multidimensional lookup tables (LUTs) are applied that transform the color from RGB to CMYK separations. Some DFEs also use trapping to mask registration errors. folding. image resolution enhancement.g.2. etc. Unlike the workstation. stapling. In some cases. An imaging module then generates a rasterized format of the PDL document at the correct resolution (e. and antialiasing) to a form cyan (C). Multidimensional.. brochures. segmentation. the input document is transformed from its PDL format such as PS or PDF. The above is usually referred to as RIP. etc. tiff. digital documents are imaged on the plate. After the proofs are approved for content.. this is done by ﬁrst utilizing an interpreter (e. and black (K) that is speciﬁcally designed and optimized for a particular digital printing system [3]. performing process-engineering functions.g.

Trapping parameters are speciﬁc to a device and its colorant set. of course. Once the RIP is complete.3 Key processing elements in the DFE and the electrophotographic (EP) process.. 1200 dpi) than what the printer is designed to handle (e..6 Control of Color Imaging Systems: Analysis and Design C Cyan TRC C΄ Halftone and resolution enhancement Halftone and resolution enhancement Halftone and resolution enhancement Halftone and resolution enhancement ESV Latent sensor image ROS M Magenta TRC M΄ Y Yellow TRC Y΄ K Black TRC K΄ MIMO control algorithm Target Clean Erase Charge Photo receptor Transfer Expose Develop Paper path FIGURE 1.g.g. 8:5 Â 11 in:). This sequence..g. part of the rendering process. Dependent on the option selected by the user. where each separation is made up of 8 b=pixel. 600 dpi) for a given paper size (e. 600 dpi) and then subsample it to the appropriate printer resolution using standard techniques.g. the input job is transformed from a PDL format to CMYK separations ready for engine consumption. Unlike offset presses. The separations are usually generated at the engine resolution (e.4 DIGITAL PRINT ENGINE (ELECTROPHOTOGRAPHIC) The print engine—sometimes referred to as the ‘‘marking engine’’—is designed to convert the electronic CMYK separations provided by the DFE into hardcopy color prints. Figure 1. digital print engine technologies are still evolving to improve . results in a slight loss of sharpness but reduces aliasing effects. 1. DFEs may RIP a given input document to a higher resolution (e..3 illustrates a typical digital press or printing system based on the principles of electrophotography (EP) invented by Chester Carlson in 1938 [4].

tandem. used in the DC8000 printer. The toned image is ‘‘transferred’’ to paper by electrostatic forces and made permanent by ‘‘fusing. and black are developed on top of the prior toner layer.1 IMAGE-ON-IMAGE AND TANDEM PRINT ENGINES Several print engine architectures have been developed to produce full-color prints image-on-image (IOI). corresponds to the desired image to be printed. This four-color image is then transferred in a single step to the substrate. it is insulating in the absence of light and conductive when light is present. The fundamental difference between tandem architecture and IOI architecture is where the four-color CMYK image is constructed. develop. expose. fuse. are then utilized to selectively expose—through the use of raster output scanners (ROS)— the charged PR drum or belt according to the binary halftoned image pattern. with outlines of the feeders. The PR is light sensitive.5). transfer. the EP process is utilized today as a key technology for high-volume full digital color printing in which four primary colors (cyan. four color process. it has its own limitations regarding volume and speed. On the other hand. marking paper handling.’’ where heat and pressure are applied to melt the toner particles and adhere them to the paper. etc. productivity. yellow. In the IOI architecture. Kodak stream ink-jet technology is a continuous system that reportedly enables offset caliber reliability. and ﬁnishing systems. yellow.’’ where a high voltage wire deposits electrons or ions on the PR in the dark causing a uniform charge buildup. detack=stripping. charge=recharge. Although ink-jet printing is another digital technology that is architecturally simple. and quality with the full beneﬁts of digital printing for high-volume commercial applications.4 shows the ‘‘skeleton’’ view of the system.An Overview of Digital Printing Systems 7 image quality. pretransfer. magenta. The toner-based digital printing process involves a circulating photoreceptor (PR) in the form of a belt or a drum. . used in the iGen3 and iGen4 production presses. each primary is ﬁrst developed on an individual PR and then separately transferred to the paper directly (or through an intermediate transfer drum or belt). and black) are developed by architecting the engine in six basic xerographic process steps: charge.4. and develop) are performed up to four times to achieve the single pass. and cleaning. and substrate latitude. In the tandem architecture (Figure 1. The CMYK separations. The ﬁrst step in the EP process is ‘‘charging. Three of these steps (charge=recharge. transfer. Once the magenta layer is developed. cost. expose. thus making it more difﬁcult to maintain high accuracy in the registration of each primary because of the four separate transfer steps. productivity. expose. the PR is mechanically and electrostatically ‘‘cleaned’’ of residual toner and then recirculated to the charging system for the next image. cyan. It is then ‘‘developed’’ by depositing oppositely charged toner particles exclusively in the charged regions thus forming a toned image on the PR. The advantages of single step transfer are that it eliminates at least four opportunities for image misregistration or disturbance or transfer efﬁciency loss. the four-color image is constructed one on top of the other on the PR belt in one complete revolution. Finally. Figure 1. called the latent image. 1. develop. [5–7]. provided by the DFE. The basic steps of the IOI xerographic process used in iGen3 and iGen4 print engines are precharge. and clean. The resulting spatial charge distribution.

Dec. to allow each individual engine to work independently of the other. L. 4.. Vol. At any time. 2003. a media transporting system. and a network of ﬂexible paper paths. of software Closed loop process controls • 5 M lines of wires • 3. depending upon the circumstances.) Multiple laser imager Fuser Paper path Black station Magenta station Cyan station Yellow station FIGURE 1.K. et al.4. which includes two printers stacked vertically.2 PARALLEL PRINTING SYSTEMS The tightly integrated parallel digital printer (TIPP) architecture developed by Xerox is designed to deliver the power of two or more printers with the simplicity of one.8 Control of Color Imaging Systems: Analysis and Design Feeder Image fully constructed on rigid belt Shuttle vacuum feeder Marker Finisher • 85 computers 600 × 4800 pixels/in. 9–11. A side view of a parallel printing system is shown in the schematic of Figure 1. (From Mestha.5 miles Automatic adjusting • 192 sensors • 102 motors decurler Standard Six servos DFA steer paper Creo to image Spire DocuSP Gripper less Wide radius paper turners handling Straight paper path Standard DFA FIGURE 1. 4096–4108. This architecture also makes the printer smart enough.5 Physical layout of the DC8000 print engine (tandem architecture).4 Physical layout of the iGen3 print engine (IOI architecture). 1. which delivers media to and from each of the printers. both printers can be simultaneously printing so that they . With permission. pp. 42nd IEEE Conference on Decision and Control. The TIPP architecture includes the software and hardware that enables multiple print engines to work together seamlessly as one printing device.6. Control elements in production printing and publishing systems: DocuColor iGen3.

The ﬁnisher may also include a purge tray for diverting sheets in order to maintain the integrity of the job.An Overview of Digital Printing Systems Printer 2 9 Control Printer 1 FIGURE 1.7) with its tightly integrated serial printing (TISP) architecture. as known in the existing art. Gray balance for a printing system of multiple marking engines.6 Schematic side view of a parallel printing system. and the like. The printers can be fed with paper from a variety of feeder modules loaded with a variety of paper types. Thereafter. At the Drupa 2008 trade show. even if the other stops. bypass pathways. page one may be printed by printer one and page two by printer two. A tandem Xerox Nuvera 288 digital perfecting system is a twin-engine monochrome press that uses a technology that keeps one engine running at full speed. Job output trays may include one or more special trays for multiple job collections. and other devices capable of marking an image on a substrate. to direct the print substrate between the highway and a selected printer or between two printers. Printers may include electrophotographic printers. For example. As such. when printing a two-page document. More than one print job can be in the course of printing at any single time instant. thus providing the user with the productivity and speed of a multi-engine printer. thermal head printers that are used in conjunction with heat-sensitive paper. including solid ink printers. ink-jet printers. The media handling system may include highways that extend from the feeder module to the ﬁnisher and pathways to transport the print media between the downstream media highways and selected ones of the printers. Thus many different printing technologies can be incorporated within a TIPP system with a complex network of media paths so that all engines work as one.) act as one. 7. Sept. reverters. The master is linked by a network of paper pathways to sensors in the printing system. interposers. (US Patent Application 20060197966. One of the printers may be designated as a reference or master printer with the remaining printer considered a slave. Xerox demonstrated the Xerox ConceptColor 220 (Figure 1. 2006. The highways and=or pathways may include inverters. these sheets are sent to ﬁnisher in the correct order. where pages one and two may be formed on opposite sides of the same sheet (duplex) or on separate sheets (simplex). A ﬁnisher with different ﬁnishing capabilities may receive printed sheets from any of the printers. It takes . multiple printers can be employed in the printing of a single print job.

there will be enhanced need for distributed optimization of interconnected workﬂows and outputs to ensure print quality consistency. TISP. Realization of enterprise-wide optimization will require substantial progress in a number of key technical areas. or cluster printing. users can achieve greater printing economics by getting twice the speed and twice the productivity with a single operator. (3) the use of sensors to measure. 1. real-time scheduling policy decisions based on sampling the current state of the press. control. Hence. These are the technologies that show promise for making the printed images appear like offset and consistent in a single job. which becomes signiﬁcant when distributed color optimization is required. They are (1) automation of the entire publishing. prepress. both at regular iGen3 speeds. but with duplex jobs 110 pages per minute. As increasing numbers of press systems within the print shop become parallel. By integrating two iGen3 engines in-line. the speed will be 110 pages per minute. production. For simplex jobs.10 Control of Color Imaging Systems: Analysis and Design FIGURE 1. with similar print engines or with loosely connected heterogeneous print engines. and creation stages and hence standardization of the color interfaces between heterogeneous press modules. the iGen3 press and doubles the effective print speed to 220 images per minute to produce 110 cut-sheet duplex pages per minute. it can approach a monthly print volume of up to 7 million color pages. which saves time and labor.5 EVOLUTION OF CONTROLS TECHNOLOGY FOR DIGITAL PRINTERS—COLOR CONTROLS VIEW Xerographic printing process used for copying and printing has evolved over many years starting from Xerox’ 914 era. and decision processes via feedback control using real-time functional press models. and between multiple machines whether intended for use in an ofﬁce or an enterprise-wise production printing and publishing systems. (2) optimization of workﬂow layouts. job to job. in this book. . Print engine 1 prints side 1 and print engine 2 prints side 2. A brief review of the history of the evolution of controls in EP products can shed some light on why controls technology is considered important in this process. and (4) the management of sensor to sensor variabilities throughout the press. A single book is not enough to cover all of the science and technology that have gone into developing these systems. and export color to prooﬁng. The ConceptColor 220 builds on the tandem architecture technology of the Xerox Nuvera 288 Digital Perfecting System. With the ConceptColor 220. This can later help us to understand the complexity of advances involved in designing digital production color printers.7 Schematic side view of a TISP parallel printing system. we concentrate on the most important imaging and control technology. as in TIPP.

electrostatic charge. and 4050 copiers helped in reducing service calls for background and density variation by more than a factor of 10 and resulted in lower subsystem costs.An Overview of Digital Printing Systems 11 In the EP process. The control functions should maintain the mass by adjusting the electrostatic charge. Due to architectural constraints and unstable charge on the PR. toner mass has to be tightly regulated so that the printer maps the desired tone to the actual output. these sensors helped to accurately set up the process and perform good diagnostics to reduce cost. the perceptible page-to-page differences were minimized. In addition to runtime controls. as the electrostatic image is developed with a charged pigmented thermoplastic powder that is transferred and fused to paper under heat and pressure. a midpoint tone . development ﬁeld. Xerox developed a low cost chargemeasuring sensor called an electrostatic voltmeter (ESV). somewhat frequent manual adjustments were made to the toner control system. called a tone reproduction curve (TRC). A two-patch control system was developed in the 1980s that controlled the toner concentration (TC).and high-density patches were created as surrogates to customer images on the PR. mass of the toner particles on the paper can vary. 5390 (1993). Since selenium alloy PRs were used. The presence of a two-patch control scheme in the Xerox 1075. which contributed to increased customer satisfaction rating. At the least. automated setups changed PR replacements from a 45 min service call to a 10 min customer operation. Similarly. This control strategy was subsequently adopted in various forms in the 5090. In 1970. which subsequently improved developability. The same sensors were also used with more advanced software in the 5100 (1991). Although dependency of various parameters to output quality was reasonably well understood by developers. 813. which appeared in the 1065 marathon copiers in 1987. With runtime controls. 4135 (1991). and the data were used in a single-input single-output (SISO) closed-loop conﬁguration to control the electrostatic charge. and 2400 copiers) to adjust these control actuators since the sensors were not reliable. and hence the developability on the PR. Low. 1090. a high-density patch was used to adjust the TC in the developer housing at a lower rate. A reduced-cost version of the infra red density (IRD) sensor was also developed. In many accounts. At that time. no closed-loop control was applied to the early copiers and printers (the Xerox’ 914. No separate charge control was required in those printers. automatic density control (ADC) sensors were introduced with Xerox 4000 duplicators and subsequently used in various forms in the 5600 and 9200 families [8]. at various stages in the printer path. The charge control was done using a separate ESV sensor followed by the developability control with the low cost IRD sensor. This is achieved by creating and implementing inverse maps in a one-dimensional (1-D) coordinate space. and the DocuPrint 4635 (1994). Using the ADC sensors. The use of different PRs and the demand for improved copy quality led to the need for better controls in the 1980s. For color printers. the coupling was removed by running the electrostatic charge control loop at a much faster rate. Low-density patches were measured by the sensor. the system was fairly stable. and transfer currents. Although high-density patch measurements are sensitive to electrostatic charge and TC. the DocuTech 135 (1990). copy quality tune up was required every 50k prints and process quality drifted due to environmental conditions.

There are many sensor-related issues to consider when the embedded spectrophotometers are located inside the print engine. uncoated. for generating the reference target (aim) measurements. A hierarchical and multilevel control loop architecture was simultaneously developed to address these problems.g. For example.. but also adjust the color for varying media conditions and wide array of media stocks (e. Color EP printers should also retain highlight. in-line embedded spectrophotometers are used in the paper path to measure ‘‘just-fused’’ toner patches. Similarly. technological advances are required in different areas including sensing. and processes. In addition to these. A correction that adjusts the sensor output to correlate between color sensors in different printers may be required to maintain instrumentto-instrument variability to be within tightly speciﬁed limits. midtone color balance without injecting any contours or blocking and offer high deﬁnition image quality at high speed. A change in temperature can cause a chromatic shift in color pigments. Neutral grays and highlights should be reproduced at offset quality. algorithms. and specialty) in order to compensate for overlay colors. the control challenges are several orders of magnitude higher since they compete in the traditional offset market that has a reputation for high quality. noninvasive measurements at highspeed. in the Pantone1 matching system. To compete. not require too much time on the press to adjust the color manually while following the international color consortium (ICC) [14] workﬂow for producing higher performance photo quality prints. the . color EP printers also need color accuracy improvements against industry standards (e. General Requirements for Applications in Commercial Offset Lithography [13] [GRACoL]. temperature. that is. To make them competitive with offset printers in terms of operational cost. such as an X-Rite iSis or DTP70 Autoscan spectrophotometer. For production quality color EP printers. many of the press makeready costs should be eliminated using automated setups. PR aging and wear in drives. due to many new challenges. Materials affect the print quality stability and stability of color balance in prints. smooth. to deliver quality and high productivity at a low run cost. Because measurement devices have different response rates. accurate color control methods require accurate and repeatable color measurements referenced to some ‘‘golden measurement standard’’ device or sensor in the master printer. humidity. textured. coated. Thus. They should maintain much tighter control on image registration between separations (for simplex and duplex functions) and paper motion at various regions in the paper path. The control loops should not only maintain process stability for individual color separations. New density sensors were developed for measuring various tone densities for monochrome and color toners. they should also closely match offset printing standards in time. Each color standard may specify a particular device.12 Control of Color Imaging Systems: Analysis and Design adjustment is required during runtime. sheet-to-sheet differences. etc. an iSis spectrophotometer.. the device-independent targets supplied by Pantone are measured by their standard instrument. International Organization for Standards [ISO]). To achieve paper=media-based contactless.g. which required the use of modern control theory and methods [9–12]. photo smoothness in faces should be retained while maintaining sharp background and shadow details. In-line spectrophotometers sensors have been found to give different measurements on justfused sheets as compared to those when the printed sheets have been cooled.

8. which becomes the foundation for navigating through the material of this book. Achieving accurate and repeatable readings from the sensor also depends on the stability of the electronics. and imaging instruments. of the color. The processing that goes on in the DFE will be at slower timescale than in the print engine.8 Hierarchical time-based color and process control functions used for improving consistency in a color EP print engine system. Sensor readings are very dependent upon the displacement of the media from the focal point of the sense head. expanding on the traditional prepress and removing limitations of printing methods. an example multilevel hierarchical structure is shown. the light source and its wavelength band. such as desktop publishing. Image capture and manipulation software. 1. A time-based separation is adopted with higher level functions occurring at a slow timescale near the prepress and faster real-time control functions typically occur in the print engine.An Overview of Digital Printing Systems 13 Image CMYK Recipe DFE Image path CMYK 3-4D LUT ICC Spot profile colors TRCs Gray balance PE Image path TRCs Level 3 Level 2 Electrostatics Level 1 Development system Process controls Color controls Targets ESV TC Loop Density sensors Inline/offline spectrophotometer FIGURE 1. the processing of images can occur at various levels inside and outside the printing and publishing system hierarchy. but at a much faster timescale than the prepress. that is. A whole segment of the printing industry is devoted to prepress. so that the measurements are closer to those of a golden measurement standard instrument. and the number of photons the system can integrate while the color samples are present underneath the sensor on a passing medium. or through special displacement insensitive (DI) optics [15]. which may be counterproductive for paper path control system. software tools. Digital prepress work is largely aided by modern computing hardware.6 PREPRESS-BASED PROCESSING Prepress refers to the preparation of digital ﬁles for printing that begins after the design decisions are made in the creation stage and ends when the document hits the press. In Figure 1. Such temperature related shifts have to be compensated for. L*. either through tight control of the media in the path. the effect of displacement of the medium on sensor output is signiﬁcant and should be compensated for.1). . Many of the processing techniques required for imaging and control functions occur at multiple levels. These are some of the factors creating error in color measurements. There is also a time-based hierarchy being adopted local to the print engine. cooling of glossy images can cause a shift in the lightness component. As pointed out before (Figure 1. As the medium moves through the point where the sensors are mounted in the paper path.

Preﬂight is a ﬁnal checklist to ensure that the ﬁles are ready for printing. the control functions are discretely handled at a lower timescale based on capturing the model of the imaging system and processing images with a myriad of algorithms. transforms. To help the reader understand how to extract spatial frequency-based models. Chapter 2 contains relevant theoretical fundamentals of important digital image processing topics such as image formation. etc. The design inspection involves looking for image quality defects (e. The transformation usually is in the form of a multidimensional LUT. up=down sampling. These transformations have to be accurate and should not induce unnecessary image artifacts. As a result. cropping. Any color that falls between is interpolated using a standard technique like trilinear or tetrahedral interpolation. in this stage. which is generated by measuring known color targets for sample of colors [16–20]. contours. Image processing techniques like denoising. Thus. the prepress environment would incorporate an accurate color model of the production environment within the design tool. In recent years. One typical function of the color management module would be to correct=compensate for scanner or camera artifacts with respect to the tone reproduction by calibrating and characterizing the devices and further compensating for device differences. which in turn helps to view the color images on a calibrated monitor (soft prooﬁng) or a proofer (hard prooﬁng). For viewing images. the multidimensional LUTs are used to transform the ﬁle ﬁrst to a deviceindependent color space and then on to the monitor color space. quantization. etc.. A good test image is useful for evaluating monitor’s quality and calibration as well as the match between the monitor and printer. Generally speaking. The spatial models could help to perform diagnostics and design inspection for production anomalies. and blue (RGB). color manipulations. They are deﬁned in an 8 bit three-dimensional (3-D) color space whose components are red. green. . color balance. and histogram modiﬁcation. smoothness. This process requires accurate characterization and calibration of the monitors prior to viewing images so that the soft-proofed images match the actual prints. image sampling. Material covered in Chapter 2 is also helpful for processing images in the DFEs. the color management module needs to transform the color from a device-dependent space (a speciﬁc scanner RGB) to a deviceindependent space (L*a*b* or independent RGB) to ensure quality color reproduction. denoising. resizing.14 Control of Color Imaging Systems: Analysis and Design color management software. image processing has become more sophisticated and more prevalent in the digital print production industry.g. are applied to the image or video frame as dictated by the user prior to its inclusion in a given document [3]. Failure to prepare all digital ﬁles can cause delays and cost overruns. They can also act as sensors to perform color analysis of the image. multimedia handling software (speech to text conversion) are used heavily. ﬁltering. deblurring. In an ideal preferred print workﬂow. Scanners and digital cameras usually capture color images in RGB type format. two different scanners imaging the same spot in a hardcopy will generate different RGB values. we also introduce the optical transfer function and modulation transfer function of imaging systems. These models can be incorporated in the printing workﬂow of the production environment. the loss of shadow or chromatic details. where each channel is quantized to 8 b=pixel.) prior to running production jobs.

gamut mapping. Normally.3. which writes the digital bits of color separated image to the PR belt. Pantone matching system). Automatic spot color editing (or control) (ASCE) function corrects with respect to device-independent L*a*b* reference values (e. The image data are then sent to the analog controls of the ROS. ASCE automatically reads print engine L*a*b* values using in-line sensors. gray component replacement (GCR)=under color removal (UCR). DFEs will provide the user with rendering control. As described in Section 1. and K images and rasterized. The image data and control information will cause the print engine to produce the best possible rendering of the user’s intent. In the DFE image path. We now discuss the press control functions by ﬁrst describing the physical print station. color jobs are separated into C. and the rest of the xerographic process follows (in the marker module) to create the printed jobs. These images are then compressed to optimize for different types of input and stored on the image disk of the DFE. tone adjustment and the multidimensional color transforms are critical control points where feedback from the internal and external paper-based spectral or color measurements are used to develop 1-D (single-channel linearization or gray balance). Accurate and consistent spot colors are also important for catalogs. compares them to the reference values.23] with an in-line spectrophotometer automatically generate multidimensional proﬁles for each halftone screen and media so that the colors match offset printing standards and the rendering is of photo quality. users can adjust tone curves manually in the DFE to modify image appearance to match their intent.g. that chapter shows how the spot color control approach can be applied to create a 1-D gray balance TRC and 2-D transforms. All of the key press control loops are resident inside the marker module. black point compensation are included in this control function. two-dimensional (2-D) or 3-D transforms [21]. Modern DFEs also perform an automated color check using in-line spectral sensors to tell the press operator if the press is ready to go into full production or if other activities are needed. Chapter 8 contains same basic ASCE algorithms. for many. In addition. Various rendering intents (or user preferences). where the jobs are actually printed on the media. Y. repeatability is as critical as or even more important than accuracy. Underlying image and signal processing and control algorithms of these control functions are described in Chapters 7 and 8. M. Often. The data are then further processed for anti-aliasing before the appropriate halftone is applied. Furthermore. business cards. These proﬁles are updated at user’s request. Advanced color proﬁling solutions offered by Xerox [22. .. and modiﬁes the CMYK recipe for each spot color to minimize the difference. This data is then transferred over high-speed image data lines to the marker module of the print engine. When the print engine requests the job ﬁles. it is decompressed in real-time off of the image disk. Customers printing applications such as marketing collaterals and direct mail can be very sensitive to spot color consistency and.7 DFE-BASED PROCESSING The role of the DFE is to convert speciﬁcations of the user’s intent into print engine data and control information. and design documents.An Overview of Digital Printing Systems 15 1. This eliminates unneeded adjustments and ensures the press is put into service as quickly as possible. multidimensional proﬁles do not accurately render process spot colors.

by linearizing the tone levels to each of the input tone values of the primaries. background. One obvious place to look for more actuators is in the image. Chapter 9 describes relevant theory and practical controller design based on state variable methods with pole-placement and regular linear quadratic design. A generic implementation is described in Refs. the number of actuators required is more than those currently available in levels 1 and 2. These constraints limit both nominal device performance and the ability to achieve that performance repeatedly. [10–12].16 Control of Color Imaging Systems: Analysis and Design 1. disturbance due to demand for toner usage (coming from each page of the job). are utilized to provide the appropriate feedback for process control. Repeatability is a function of the process control system used in the print engine. The amount of toner mass deposited on the PR is measured at different tone levels using calibrated optical sensors.8 PRINT ENGINE-BASED PROCESSING The print engine and the DFE are responsible for most of the constraints regarding image quality. In addition to the real-time process adjustments. which contains time-based hierarchy which is architecturally named as levels 1. density (or dot gain). The control algorithms employed to control the process are often customized for the underlying system architecture to achieve optimum stability results. and color stability. and 3 controls [11] and not associated with any particular print job. color balance. Nominal performance is a function of engineering trade-offs in the design process. actuator saturation. only individual separations are controlled. For levels 1. In real-time. since electronically produced color documents contain pixels that are described in a 3-D RGB space. we can achieve improved controls for all separations individually [21]. 2. before being sent to the printer. This is perhaps one of the most complex digital SISO control loops to analyze which comprises of unstable time varying plant with time delays. So. as in a typical inventory management system. depending on the sensing method. This type of tone adjustment is called level 3 controls and can be performed on the PR belt or on the paper. is maintained to some desired set point for each of the color station. These are called level 2 control loops. which is the ratio of toner mass to carrier plus toner mass. and 3. to control primary color mixtures for optimum color quality. Surrogate patches. We also show how observers can be used to compensate for time delays in the TC control system. sensor noise. Charge control loop adjusts the PR charge and the intensity of the laser in a level 1 subsystem loop as indicated in Figure 1. This information is then used to control the dot gain and development reproduction curves by actuating the charging and development system actuators. These color pixels are transformed to corresponding digital CMYK values to a printable form. 2. For example.7 so that the voltages on the PRs maintain or track the desired values within a small tolerance (generally less that 1%) to prevent the appearance of unwanted variations in prints. and a variable actuation cycle. placed in the interdocument zone (IDZ) in between images. . This is accomplished with digital controllers in the developer housing using TC sensors and actuating the dispense rate. and developed tone reproduction control functions for each of the separations and provide information for online remote interactive diagnostics. TC. PR voltages are read using an ESV sensor. controls of this nature run at a much faster rate and include charge.

Spatial color correction is required in EP printing process [24] due to the streaks and bands that are inherent to the print process. In addition to the opportunities provided by these models to create robust control and imaging system. pp. in 42nd IEEE Conference on Decision and Control. 9–12. BF Kuvin. JJ Folkins. and PY Li. 31–Nov. Vol. mathematical tools for the analysis and design of open. process. 37–39. and methods required for developing and optimizing a digital printing system and producing state-of-the-art color quality. Therefore. efﬁcient use of multidimensional interpolation algorithms. Control elements in production printing and publishing systems: DocuColor iGen3. 2006.. US Patent 5. 5. LK Mestha. Jul. Chapter 10 presents actual physical models of the EP process with access to principal image. 6. 8. These two chapters also contain necessary information for characterizing printers. and methods to generate good color transformations. LK Mestha et al.gov 2.and closed-loop control systems are required to optimize an internal press control system. 2005. Transfer. pp. US Department of Labor. Nov. Modular press control. and marking subsystem parameters that simulate the development of fused prints. Dec. 323–327. 2003. They are brieﬂy mentioned in this book. 4. 1996. www. in The Optics Encyclopedia. Mar. 4096–4108. JJ Folkins. 2-. 5. 19. Color gamuts for different settings of printer parameters can be generated using these physical models. 2003. Five cycle image on image printing architecture. REFERENCES 1. E Saber. R Lux and H-J Yuh. Nov 2002. IEEE Signal Processing Magazine. to provide an accurate system inverse. CMYK separations).824.576. Chapters 3 through 5 provide a strong mathematical foundation to design modern multivariable discrete control systems. Publishing. . cleaning and imaging stations spaced within an interdocument. Bureau of Labor Statistics. May 5.g. Chapters 7 and 8 describe the use of image actuators (e. US Patent 5. DSP utilization in digital color printing. 3. 1998. UT. or 3-D spatial model structure for the development of component primaries and then to mixed colors on the paper are some of the new developments that can enable spatial image quality compensation [25]. mapping out-of-gamut colors to the surface. algorithms.An Overview of Digital Printing Systems 17 In recapitulation. Wiley-VCH. which are required for improving the press stability using process actuators. Career Guide to Industries: Printing. Electrophotographic copying and printing (xerography). 2004. S Dianat. A dynamic 1-. Berlin.. Oct.576. they can be extended to control the color by rendering spatially color corrected pixels. this book includes a collection of theoretical techniques. 4. DA Hays and KR Ossman. In addition to the fundamental knowledge in image processing and image transforms.bls. pp. Salt Lake City. Is image-on-image color printing a privileged printing architecture for production digital printing applications? NIP20: Proceedings of the IS&T’s International Conference on Digital Printing Technologies. Chapter 6 describes the multidimensional interpolation and smoothing=ﬁltering techniques useful for generating inverse maps. Metalforming. 7. largely in the DFEs and to some extent in the prepress.824.

918. Boston. 5.. US Patent 6.380. New Xerox color press delivers breakthrough image quality: Drives more proﬁtability from digital printing. Xerox News Release. 2004.902. 2006. 1:2004-10 (Proﬁle version 4.481. Trieste. and LK Mestha. Xerox News Release. 25. A full GRACoL Technical Speciﬁcation document. Dusseldorf. Method and system for providing closed loop color control between a scanned color image and the output of a color printer. UT.0. Mar. 2004. LK Mestha. Control advances in production printing and publishing systems. 2002.551. 1–4. 6.E. 18.721. 1997 (contributions by R. LK Mestha et al. Method and apparatus for calibration of color values. proﬁle format. 31–Nov. 1998. International Color Consortium Speciﬁcation.809. Oct. May 7. Colorimeter and calibration system. 15.. MA. 18–22. 21. US Patent 5. Germany.384.612.853. Image technology colour management—Architecture. and R Bala. 2008. 2. Spectrophotometer for color printer color control with displacement insensitive optics. 24. Apr. 17. 22. SA Dianat. Sep. NIP20: Proceedings of the IS&T’s The International Conference on Digital Printing Technologies. US Patent 7. 6. Jan 2.332. and data structure.692. 14. 20.18 Control of Color Imaging Systems: Analysis and Design 9. LK Mestha and ER Viturro.525.347. Jun. CS Chan. 19. 1996. US Patent 6. Method for spatial color calibration using hybrid sensing systems.603. CB Duke et al.107. Dec. US Patent 5. MD. 20080037069. V4.. A multilevel modular control architecture for image reproduction. 13. M Stokes.750. Rochester. KD Vincent. A control-oriented survey of xerographic systems: Basic concepts to new frontiers.855. Systems and methods for compensating for streaks in images.0). 21. LK Mestha. 1994.518.910. Color System Integration. PK Gurram.271. US Patent 5. 1993. May 22. NY. Grace). NIP21: Proceedings of the IS&T’s 21st International Conference on Digital Printing Technologies. May 29. 2004. Sep. US Patent 5.633. FF Hubble III and JA Kubby. 23. 6. 15. LK Mestha. 12. 7. 25. Italy. Feb. ICC. Salt Lake City. 2005. 14.442. 2008. The best gets better: Xerox elevates high performance and image quality of iGen3 digital press. HA Mizes. 1992. 21. 16. Other US Patents 6. 7. . Sep. Calibrating. 13. Method and system for analytic generation of multidimensional color lookup tables. US Patent Application.259. R Bala. Use of spectral sensors for automatic media identiﬁcation and improved scanner correction. 2008. 10. in Proceedings of the American Control Conference.272. printing and prooﬁng to the G7 method. Baltimore.2. ES Hamby et al.382. 2-D and 3-D printer calibration algorithms with printer drift. G Bestmann. Comparison of 1-D. 30–Jul. in Proceedings of the IEEE International Conference on Control Applications. 11. Mar. Jun. 2008.

2 brieﬂy covers image formation and systems. l1 . y À l2 ) dl1 dl2 (2:1) 19 .2 DIGITAL IMAGE FORMATION AND SYSTEMS Linear system theory provides a powerful tool for the modeling and analysis of various imaging systems [1–3]. Section 2. Finally. y) to a given input f (x. then the input aI1 þ bI2 results in the output aO1 þ bO2 for any I1 .8. printers. Issues such as image resizing and its practical implementation are addressed in Section 2. Unfortunately. ﬁrst the input is broken up into sum of weighted impulses (points) 1 ð 1 ð f (x. l2 ). and the input I2 to the system results in the output O2 . A linear system is characterized as a system that obeys the superposition principle. I2 signal and scale factors a and b. Section 2. 2.7. including topics such as digital image formation. l2 ) in the input image plane as shown in Figure 2. ﬁltering. image sampling. referred to as the point spread function (PSF) of the imaging system that speciﬁes the output of the system when the input is a point (impulse) at location (l1 .1 INTRODUCTION This chapter is devoted to providing an overview of fundamentals of digital imaging. and image ﬁltering is covered in Section 2. Image enhancement is covered in Section 2. l2 )d(x À l1 . Section 2.4 discusses image sampling and quantization. Image degradation and restoration are brieﬂy discussed in Section 2. and the human visual system (HVS) are among many examples of imaging systems that are modeled and analyzed by using linear system theory. if the input I1 to a system results in the output O1 . y. quantization. imaging systems. that is.3 covers optical and modulation transfer functions. scanners.6.1. such systems are almost always approximated by linear systems to make their analysis mathematically tractable. basic image halftoning techniques are described in Section 2. A linear system provides a convenient model for an imaging system.10.2 Fundamentals of Digital Image Processing 2. To ﬁnd the output of an imaging system g(x. However. A two-dimensional (2-D) linear imaging system is characterized by a function h(x. y) ¼ À1 À1 f (l1 . y).5 deals with image transforms. Conventional and digital cameras. and image transformation. none of the imaging systems encountered in the real world are completely linear. Section 2.9.

1 PSF of an imaging system. It assumes that the shape of the PSF depends on its location in the input image plane.1 POINT SPREAD FUNCTION OF A DEFOCUSED LENS A simple example of an LSI imaging system is a defocused lens. 1 ð 1 ð g(x. l2 )h(x. or capturing a picture of a scene with objects at various planes. l2 ) ¼ h(x À l1 .20 Control of Color Imaging Systems: Analysis and Design h(x. that is. y. y) ¼ À1 À1 f (l1 . Next. the output image is determined as the sum (integral) of the outputs of all the impulses that make up the input. y À l2 )dl1 dl2 (2:5) A major advantage of LSI systems is that they can be analyzed using Fourier transform theory. l2 ) (2:2) Finally. y À l2 ) (2:4) As a result of this.3 is known as Fredholm equation of ﬁrst kind. Now. λ1. y. y–λ2) λ2 x λ1 Imaging system y x λ1 FIGURE 2. the PSF can be characterized as a function with two arguments x and y. y. the output for each weighted impulse is determined from the impulse response f (l1 . y À l2 ) ! f (l1 . using the linearity principle. or simply h(x. An exact method of ﬁnding the PSF of such a system is based on physical optics and has been described . l2 )h(x À l1 . many imaging systems are approximated by LSI systems. In many instances. For that purpose. The integral equation given by Equation 2. y). Such a system is known as a linear space-invariant (LSI) system. The input–output relationship simpliﬁes for such a linear system and is called convolution integral. λ2) y λ2 δ(x–λ1.2. Examples for such system are the subject motion when a picture is being captured with a stationary camera. l2 )dl1 dl2 (2:3) This is a general input–output relationship for a linear imaging system. l1 . 2. y. we give simple examples of LSI imaging systems. Such a system is referred to as a linear space-variant system. l2 )h(x. 1 ð 1 ð g(x. y) ¼ À1 À1 f (l1 . the shape of the PSF is independent of its location and is thus the same everywhere. l2 )d(x À l1 . l1 . l1 . h(x.

which is moving with a speed of v in the horizontal direction. that is. y)dl 0 (2:9) where * stands for convolution.Fundamentals of Digital Image Processing 21 in detail in a paper by Lee [4]. y) ¼ x0 x0 ð f (x À l. y) ¼ 1 T T ð f (x À x0 (t). y) * h(x. we can rewrite the above integral as 1 g(x. Thus. Using a change of variable l ¼ vt. y). y) ¼ &1 x0 d(y) 0 x x0 0 otherwise (2:8) Consequently. the 2-D PSF is a line of length x0 . R is the blur radius and is given by R¼ f 2 jL À L0 j 2F LL0 (2:7) where f is the focal length of the lens F is the f number of the lens that is the focal length divided by diameter of the lens aperture L0 is the distance at which the lens has been focused L is the distance from the object to the lens 2. The PSF is given by & 1 x2 þ y2 R2 2 (2:6) h(x. the image g(x. h(x.2 POINT SPREAD FUNCTION OF MOTION BLUR Another example of a linear imaging system is the motion blur resulting from camera motion or object motion. y) ¼ T T ð f (x À vt. a point on the object moves by a distance x0 ¼ vT. y À y0 (t))dt 0 (2:11) . An alternative approach based on geometrical optics results in a reasonable approximation of the 2-D PSF as a circle of unit area with a radius R. During the exposure time T. Consider an object with uniform motion in the horizontal direction characterized by a constant speed v. is found by the following equation: 1 g(x. y) of an object f (x. y) ¼ f (x. y)dt 0 (2:10) The above equation can be generalized for any type of motion as g(x. y) ¼ pR 0 otherwise In this equation.2.

where a is acceleration in the x-direction. vy ) ¼ H(0.and y-directions x0 (t) ¼ vx t and y0 (t) ¼ vy t. printer. Finally. This will be described in detail in Chapter 3. the OTF is one-dimensional (1-D) since there is no motion in the y-direction. then x0 (t) ¼ 0:5at 2 þ vx t. and the lens.3 OPTICAL AND MODULATION TRANSFER FUNCTIONS The optical transfer function (OTF) of an imaging system is deﬁned as the 2-D Fourier transform of its 2-D PSF and is given by the following integral 1 ð 1 ð H(vx . consider uniform motion in the x-direction. that is. The actual PSF of the eye is the cascade of these two PSFs [3. for uniform motion in both x.22 Control of Color Imaging Systems: Analysis and Design where x0 (t) and y0 (t) are the equations for motion in the x. constitute a blur type (low-pass) PSF. 0) (2:13) As an example. If there is acceleration in the x-direction. y)eÀj(vx xþvy y) dxdy (2:12) Either the OTF or the PSF completely characterizes an LSI imaging system. vy ) ¼ À1 À1 h(x. The OTF of an imaging system is a complex quantity and consists of magnitude and phase. and display device is the convolution of the PSFs of the three imaging subsystems. The modulus of the OTF normalized to a value of 1 at zero frequency is referred to as the modulation transfer function (MTF). vy ) M(vx . 2. This 1-D OTF is given by 1 OTF(vx ) ¼ x0 The corresponding MTF is À Á OTF(vx ) sin vx2x0 ¼ MTF(vx ) ¼ OTF(0) vx2x0 Figure 2. For example. that is. while the rod and cone detectors’ inhibitory response constitute a sharpening type (high-pass) linear system.3 POINT SPREAD FUNCTION OF HUMAN VISUAL SYSTEM The third example of an LSI imaging system is the HVS.2 shows MTF of uniform motion in the x-direction.2. H(vx . the cornea. the PSF of a printing system consisting of a scanner. The optics of the eye.5]. respectively. (2:15) x0 ð jvx x e 0 À Á e jvx x0 À 1 sin vx2x0 Àjvx x0 dx ¼ ¼ v x x0 e 2 jvx x0 2 (2:14) . the pupil.and y-directions. 2.

vy ) ¼ À1 À1 1 ð 1 ð h(x.2 MTF of uniform motion in x-direction. consider 1 ð 1 ð H(vx . the previous analysis can be simpliﬁed.3 0.17 can be rewritten as 2p 1 ð ð H(vx . To show this.Fundamentals of Digital Image Processing 1 0. such as the lens and the optics of a digital camera. The OTF of this system like its PSF is circularly symmetric. printer.8 0. Equation 2. In such cases.1 0 0 1 2 3 4 6 5 ωx/πx0 7 8 9 10 23 FIGURE 2.9 0.4 0. The previous discussion is fairly general in the sense that it applies to imaging systems with nonsymmetrical PSF too. Now. in many practical systems. vy ) ¼ 0 0 h(r)eÀj(vx r cos uþvy r sin u) rdrdu (2:18) . The PSF of such a system is 1-D and is given by pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ h(r) ¼ h x2 þ y 2 (2:16) where r is the radial distance.2 0.6 0. we consider a circularly symmetric imaging system.5 0. or scanner. However. the PSF is rotationally invariant or circularly symmetric. y)eÀj(vx xþvy y) dxdy pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ x2 þ y2 ÞeÀj(vx xþvy y) dxdy ¼ À1 À1 hð (2:17) Using polar coordinates.7 OTF (ωx) 0.

Hence. we set f ¼ 0. 1 ð 2p ð H(vx . Thus. the OTF of a rotationally symmetric PSF is also rotationally symmetric and is related to its 1-D PSF through 1-D Hankel transform. vy ) ¼ 0 rh(r) 0 eÀjrvr cosu dudr (2:21) The inner integral is given by 2p ð eÀjrvr cosu du ¼ 2pJ0 (rvr ) (2:22) 0 where J0 is the Bessel function of ﬁrst kind and order 0. The inverse Hankel transform can be used to express the PSF in terms of OTF. Therefore. the inner integral is independent of f. 1 ð H(vr ) ¼ 2p 0 rh(r)J0 (rvr )dr (2:23) This is known as Hankel transform. This is given by 1 h(r) ¼ 2p 1 ð vr H(vr )J0 (rvr )dvr 0 (2:24) . vy ) ¼ 0 0 h(r)e rdrdu ¼ 0 rh(r) 0 eÀjrvr cos(uÀf) dudr (2:20) Since function cos (u) is periodic with period of 2p and integration is over one period.24 Control of Color Imaging Systems: Analysis and Design Using the trigonometric identity vx cos u þ vy sin u ¼ vr cos (u À w) where vr ¼ and w ¼ tanÀ1 we have 2p 1 ð ð Àjrvr cos(uÀw) 1 ð 2p ð (2:19) qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ v2 þ v 2 x y vy vx H(vx .

9 0. we have dx 2 H(vr ) ¼ 2 2 vr R The MTF is given by H(vr ) ¼ 2J1 (Rvr ) MTF(vr ) ¼ Rv H(0) r (2:27) Rvr ð 0 J1 (Rvr ) dxJ1 (x) 2 Rv dx ¼ 2 2 xJ1 (x)j0 r ¼ 2 Rvr dx vr R (2:26) where J1 (x) is a Bessel function of the ﬁrst kind of order 1 and its Taylor series expansion is J1 (x) ¼ 1 X i¼0 (À1)i x2iþ1 22iþ1 i!(i þ 1)! (2:28) Figure 2.3 MTF of defocused lens.3 0. consider a defocused lens.2 0.8 0.Fundamentals of Digital Image Processing 25 As an example of a circularly symmetric OTF. 1 0.1 0 0 5 Rωρ 10 15 FIGURE 2. The OTF is given by Hankel transform as 1 ð H(vr ) ¼ 2p 0 1 2 rh(r)J0 (rvr )dr ¼ 2p r 2 J0 (rvr )dr ¼ 2 2 pR vr R 0 R ð Rvr ð xJ0 (x)dx (2:25) 0 Using the known relationship xJ0 (x) ¼ d[xJ1 (x)].4 0.5 0.6 0.7 MTF (ωρ) 0. .3 shows the MTF of a defocused lens.

In practice. vy ) ¼ MTF x y v0 (2:29) In image processing applications. cycles/° 50 FIGURE 2. In a discrete system. The ratio of the output to input modulation gives the MTF at that particular frequency and orientation. An approach that is based on square-wave analysis is more efﬁcient. ! b qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ v r À vr v e 0 v2 þ v2 ¼ MTF(vr ) ¼ A a þ MTF(vx . a single measurement can determine several MTF points. several square waves with . The plot of MTF is shown in Figure 2. The number of pixels contained in each cycle of the square wave (which deﬁnes the fundamental frequency or the ﬁrst harmonic of the wave) would depend on the resolution needed for the MTF measurement.4 0. the following parameter values have been used: A ¼ 2:6.8 Contrast sensitivity 0.5 0. we consider the MTF of the HVS.4. As a ﬁnal example. The measurement on human subjects has conﬁrmed that the MTF of the HVS can be approximated by the empirical function given by [3. many such measurements are needed before a good approximation to the MTF curve can be constructed. Since the harmonics of a square wave contain signiﬁcant energy at high frequencies.9 0. The main problem with this approach is that even for a given orientation. v0 ¼ 8:772 cycles=degree and b ¼ 1:1. To ﬁnd the MTF of such a linear system.6 0. an input level is chosen and a square wave with two amplitudes near that input level is constructed.1 0 0 Control of Color Imaging Systems: Analysis and Design 10 40 20 30 Spatial frequency.7 0.2 0.4 MTF of the HVS. The MTF of LSI discrete systems (such as a digital printer) can be measured using DFT concepts.5]. a ¼ 0:0192. a sampled input (an image described as pixels) results in an output (printed image) that is the convolution of the input with the system PSF (in this case resulting from the inefﬁciencies of the development process).26 1 0.3 0. Using linear system theory concepts. the MTF of the system can be measured by applying a sampled sinusoid with a certain frequency to the system and measuring its output.

the output frequency will be converted to cycles=mm. for a printer at 300 pixels=in.g. In summary.g. FIGURE 2. for a printer with that resolution) and its DFT is measured. Figure 2. MTF at each discrete frequency is measured as the magnitude of the DFT of the output divided by the magnitude of the DFT of the input at that frequency.5 Square-wave target patches. a square wave with 50 pixels=cycle will have a fundamental frequency of 0. 2. To make the measurement more robust. The output image is scanned (e. the following steps are followed: 1.Fundamentals of Digital Image Processing 27 various fundamental frequencies are applied and the results of the measurements are ﬁtted to a curve. . For example. When the target is printed with a printer. to measure the MTF of an imaging system (e. 300 pixels=in. a digital printer).24 cycles=mm. (12 pixels=mm). Square-wave target with a given fundamental frequency is applied to the system (note that the DFT of a square wave does not contain any even harmonics). the fundamental frequency is 0.. for example.02 cycles=pixel. 3. The MTF is normalized to 1 at zero frequency.. this process is repeated for several square waves with different fundamental frequencies and the results are ﬁtted to a curve.5 shows square-wave target patches for MTF measurement.

they can be approximated by bandlimited functions.28 Control of Color Imaging Systems: Analysis and Design 2. vy ).4 IMAGE SAMPLING AND QUANTIZATION To process an image digitally (using a computer). vy ) is zero outside a bounded region in the frequency plane (vx . as shown in Figure 2.7) with spacing of Dx and Dy.5]. Real-world images are not truly band limited. To understand the process of image sampling. that is. …. y) is said to be band limited if its Fourier transform F(vx . The ideal rectangular image sampler is a 2-D array of discrete delta functions placed on a rectangular grid (Figure 2. we need to revisit the 2-D sampling theorem [3.6 Region deﬁning a band-limited signal. y) Δx Δy ….7 Sampling function s(x. y). x FIGURE 2. A 2-D function f (x. . it must be available in digital form. s(x.6. ωy ωy 0 ωx ωx0 FIGURE 2. however. y …. Assume that we have a band-limited 2-D signal (such as an image). The digitized image is obtained by sampling the continuous tone image on a discrete grid and quantizing the sample values using a ﬁnite number of bits.

y À mDy ) (2:31) n¼À1 m¼À1 The spectrum of the sampled signal is obtained by taking the Fourier transform of f *(x. y) f * (x. vy )eÀj(vx nDx þvy mDy ) dvx dvy À1 À1 .y¼mDy ¼ 1 4p2 1 ð 1 ð F(vx . y) s(x. y)jx¼nDx . m) ¼ fa (x. y À mDy )eÀj(vx xþvy y) dxdy (2:32) n¼À1 m¼À1 1 X 1 X f (n. vy ) ¼ À1 À1 1 X f *(x. 1 X 1 X f *(x. That is. mDy ) À1 À1 d(x À nDx . mDy ). 1 X 1 X s(x. y). y À mDy) (2:30) n¼À1 m¼À1 1 1 The sampling frequencies in the x. m)eÀj(vx nDx þvy mDy ) n¼À1 m¼À1 where f (n. then we have 1 ¼ 2 4p 1 ð 1 ð f (n. respectively. y) and s(x. Since fa (x.8. m) ¼ fa (nDx .8 Sampling process.and y-directions are fsx ¼ Dx and fsy ¼ Dy . y). vy ). as shown in Figure 2.Fundamentals of Digital Image Processing 29 fa (x. y) ¼ d(x À nDx. y)eÀj(vx xþvy y) dxdy 1 X 1 ð 1 ð ¼ ¼ fa (nDx . y) ¼ fa (nDx . y) FIGURE 2. The sampling process is basically the multiplication of fa (x. vy )eÀj(vx xþvy y) dvx dvy jx¼nDx . y) can be obtained by taking the inverse Fourier transform of F(vx . mDy )d(x À nDx .y¼mDy (2:33) À1 À1 F(vx . 1 ð 1 ð F*(vx .

30 Control of Color Imaging Systems: Analysis and Design Substituting Equation 2.33 into Equation 2. the Fourier transform of the sampled signal is the periodic extension of the Fourier transform of the analog signal with period given by (vsx . v) À1 À1 1 1 ð ð 1 X # e Àj(uÀvx )nDx Àj(vÀvy )mDy dudv n¼À1 m¼À1 1 X m¼À1 F(u. the original signal can be reconstructed by low-pass ﬁltering the sampled signal. provided that there is no overlap between the spectrum components in frequency domain or in other words there is no aliasing.32 yields F*(vx . As a result of this.vy þ v0 d[u0 À n]d[v0 À m] du dv Dx Dy Dx Dy n¼À1 m¼À1 ¼ fsx fsy 1 1 X X n¼À1 m¼À1 À1 À1 F(vx þ n2pfsx . v) À1 À1 d ! (v À vy )Dy À m dudv Âd 2p Let u0 ¼ (uÀvx )Dx and v0 ¼ 2p 1 F*(vx . Equation 2. . vy ) ¼ 1 X 1 X 2 1 4 4p2 1 ð 1 ð 3 F(u. then 1 1 ð ð 1 1 1 1 X X 2p 2p ¼ F vx þ u0 . This is possible only if the sampling frequencies satisfy.9. vsy ) as shown in Figure 2. v)eÀj(unDx þvmDy ) dudv5eÀj(vx nDx þvy mDy ) 1 X n¼À1 m¼À1 1 ¼ 2 4p ¼ 1 4p2 1 ð 1 ð À1 À1 " F(u. v) À1 À1 eÀj(vÀvy )mDy 1 X n¼À1 eÀj(uÀvx )nDx dudv (2:34) We now use the well-known identity 1 X k¼À1 eÀj2pxk ¼ 1 X k¼À1 d(x À k) (2:35) Then.vy þ m2pfsy ) (2:37) Therefore.34 can be reduced to 1 4p2 1 ð 1 ð 1 X 1 X F*(vx . vy ) ¼ F(u.vy þ v0 d[u0 À n]d[v0 À m]du0 dv0 Dx Dy n¼À1 m¼À1 Dx Dy À1 À1 X X 1 1 2p 2p 2p 2p 0 0 F vx þ u0 .vy ) ¼ 2 4p 1 ð 1 ð n¼À1 m¼À1 (u À vx )Dx Àn 2p ! (2:36) (vÀvy )Dy 2p .

any sampling rate will capture all the information contained in frequencies below one-half of the sampling rate.1 TWO-DIMENSIONAL SAMPLING THEOREM Suppose a 2-D function fa (x. To avoid aliasing. Aliasing causes the high frequencies to appear as low frequencies in the sampled signal. p vy 0 units apart. that is. y) is band limited to (vx0 . If a signal is undersampled (i.e. if a certain sinusoid contained in the original signal needs to be accurately reproduced after sampling. Dy) ¼ .and y-directions. vy0 ). sampled at a rate less than the Nyquist rate). y) is completely determined by samples of the p vx0 function spaced (Dx.Fundamentals of Digital Image Processing ωy 31 ωsy ωy0 –ωsx ωx0 ωsx ωx –ωsy FIGURE 2. Consequently.. the Fourier transform of the function is zero for all frequency pairs outside the support shown in Figure 2. Thus.9 Fourier transform of the sampled signal. it has to be sampled at least twice at each cycle. v ! 2vx0 f ! 2fx0 vsx À vx0 ! vx0 ! sx ! sx vsy À vy0 ! vy0 vsy ! 2vy0 fsy ! 2fy0 Thus we obtain the Nyquist theorem for sampling.4.6. Dx Dy 1 2fx0 1 2fy0 (2:38) (2:39) 2. aliasing will occur. 2fy0 ) samples per unit length in the x. This means that the signal has to . Then the function fa (x. This minimum required rate is called the Nyquist rate or the Nyquist frequency. be sampled at the minimum rate of (2fx0 .

The Nyquist rate for this signal is fNyquist ¼ 2f0 ¼ 2 Â 200 ¼ 400. we obtain 400p 4p 4p 2p n ¼ 2 cos n ¼ 2 cos 2pn À n ¼ 2 cos n f (n) ¼ 2 cos 300 3 3 3 Reconstructing the signal from these samples will result in a signal of lower frequency. Example 2. Once the image has been inadequately sampled and the high frequencies have been aliased into the low frequencies. the low-pass ﬁlter should ideally eliminate all the frequencies that are higher than half the sampling rate. y) Anti-aliasing filter Sampler f (n. As is obvious from this ﬁgure. there is no way to eliminate the aliasing effect. This lower frequency is given by v 1 ¼ fs 2p ¼ 200p 3 Figure 2.32 Control of Color Imaging Systems: Analysis and Design fa(x.2 Consider the analog signal fa (x.1 In the ﬁrst example. The block diagram of a sampler with an anti-aliasing ﬁlter is shown in Figure 2. The anti-aliasing ﬁlter should be placed somewhere between the lens and the sensor. the original signal should be low-pass ﬁltered prior to sampling. The low-pass ﬁlter used to reduce or eliminate aliasing is referred to as the anti-aliasing ﬁlter.11 shows the original analog signal and the reconstructed signal. This is usually due to the fact that the camera lens is of high quality and is capable of accurately reproducing scenes containing frequencies higher than half the Nyquist frequency (determined by the sensor resolution). Aliasing is a common problem in digital cameras that use CCD sensor arrays. It should be noted that the anti-aliasing ﬁltering has to be applied prior to the sampling. In particular. If we undersample this signal at the rate of fs ¼ 300 samples per unit length. y) ¼ 2 cos (400px þ 600py) .10 Anti-aliasing ﬁlter. we consider the 1-D analog signal fa (x) ¼ 2 cos (400px) The signal has one frequency component f0 ¼ 200 cycles per unit length. the signal is aliased due to undersampling. Example 2.10. m) Bandlimited analog signal FIGURE 2.

y) ¼ 2 cos Â 300x þ Â 400y ¼ 2 cos (200px þ 200py) 3 2 This means that the reconstructed signal appears as a low-frequency signal. Then the sampled signal is 400p 600p 4p 3p f (n.12a by a factor of 4 and then upsample it to the original size without using any anti-aliasing ﬁlter. Thus the Nyquist rate for the signal is fNyquist ¼ 2 Â (200.and y-directions.11 One-dimensional example of aliasing. 300) ¼ (400. respectively. . The signal has frequency components of 200 and 300 cycles per unit length in the x. Now suppose we sample the signal at a rate below the Nyquist frequency. To avoid aliasing. we ﬁrst downsample the monochromic LENA image shown in Figure 2.3 To show aliasing in an image. the sampling rate must exceed the Nyquist rate. The resulting image is aliased as shown in Figure 2. Example 2. respectively. say 300 and 400 samples per unit length in the x. the resulting continuous signal would be 2p p g(x.12b. m) ¼ 2 cos nþ m ¼ 2 cos nþ m 300 400 3 2 4p 3p 2p p ¼ 2 cos n À 2pn þ m À 2pm ¼ 2 cos À nÀ m 3 2 3 2 2p p ¼ 2 cos nþ m 3 2 Now.and y-directions. if we send the sampled signal through an ideal digital to analog converter (D=A).Fundamentals of Digital Image Processing 33 fa(x): Original analog signal Reconstructed signal = Signal samples 1 fs 1 300 Δx = = Δx 2Δx 3Δx 4Δx 5Δx 6Δx x FIGURE 2. 600) cycles per unit length.

2. 2.12 (a) Original LENA Image (b) aliased LENA Image. . the quantization decision levels are of equal length D. The function Q(x) is shown in Figure 2. 2. L where L ¼ 2B (2:40) di ’s are called decision levels and ri ’s are reconstruction levels. . the amplitude of the sampled values is quantized (digitized). There are two ways to quantize a signal: scalar quantization (SQ).2 IMAGE QUANTIZATION Once the signal has been sampled (discretized). The concept of SQ for 1-D signals is shown in Figure 2. where each signal sample is individually quantized.13 SQ for 1-D signals. .4. A quantizer is a many-to-one mapping that maps a range of input values into a single output value. where a block of signal samples is jointly quantized.34 Control of Color Imaging Systems: Analysis and Design (a) (b) FIGURE 2.1 Uniform Quantization In uniform SQ. A typical B-bit scalar quantizer for a signal with a dynamic range of xmin to xmax is deﬁned as Q(x) ¼ ri if diÀ1 < x < di i ¼ 1. The quantization step size is related to the signal dynamic range [xmin xmax ] and number of quantization bits B by f (x) f (n) f Q(n) x x FIGURE 2. . referred to as the quantization step size. .4. and vector quantization (VQ).13. 2.14.

d4 ¼ 1 r1 ¼ 0:125.14 Quantization function Q(x).2 Signal-to-Quantization Noise Ratio (SQNR) Let D denote the quantization step size of a uniform quantizer and let’s assume that the signal distribution (histogram) over a given quantizer decision interval can be approximated by a uniform probability density function (PDF) as shown in . ri ¼ di þ diÀ1 2 1 i L (2:43) Example 2.4. r3 ¼ 0:625.Fundamentals of Digital Image Processing Q(x) 35 rL r3 r2 r1 x d0 d1 d2 d3 dL–1 dL FIGURE 2.2. SOLUTION The decision boundaries and reconstruction levels are d0 ¼ 0. D¼ xmax À xmin xmax À xmin ¼ L 2B (2:41) The decision boundaries are related to the quantization step size D by di À diÀ1 ¼ D 1 i L (2:42) The reconstruction levels are chosen to be halfway between the decision boundaries in order to minimize the mean-squared error (MSE). d1 ¼ 0:25. d3 ¼ 0:75. r2 ¼ 0:375. r4 ¼ 0:875 2.4 Design a uniform 2 bit (4-level) quantizer for a signal with dynamic range between 0 and 1. d2 ¼ 0:5.

the reconstruction and . 2. that is.36 Control of Color Imaging Systems: Analysis and Design P(e) 1 Δ e – Δ 2 Δ 2 FIGURE 2. If p(x) is uniformly distributed between xmin and xmax . if the PDF of X is nonuniform. Mathematically. one would expect that in order to minimize the average error the quantizer structure should be ﬁner around the peaks of the signal distribution and coarser in regions where the signal occurs infrequently. In general. 2 2 The quantization noise power is the variance of this uniform distribution. the quantization error for any signal sample would be uniformly distributed between À D and D. When the PDF of the input signal is not available. the SQNR increases by 6 dB for each additional quantization bit. the optimum quantizer will also be nonuniformly spaced.15 PDF of quantization noise. Figure 2.2. the smoothed data histogram can be used as an estimate of the PDF.15. In such a case. 1 ð 2 ð D s2 ¼ e À1 e2 P(e)de ¼ ÀD 2 e2 1 D2 de ¼ 12 D (2:44) A B-bit quantizer for a signal in the range of [xmin xmax ] has a step size of D¼ xmax À xmin 2B (2:45) The signal to quantization noise power ratio (SQNR) is given by SQNR ¼ The SQNR in decibels (dB) is (SQNR)dB 2 À Á s ¼ 10 log s ¼ 6B þ 10 log10 12s2 À 20 log10 (xmax À xmin ) (2:47) s s2 n s2 12s2 12s2 22B s s ¼ 2s ¼ s2 (xmax À xmin )2 D n (2:46) Therefore.3 Optimum Minimum Mean-Square Error Quantizer Assume that signal X has a PDF p(x). then the optimum quantizer in the minimum mean-square error (MMSE) sense is a uniform quantizer.4. However.

2. in most cases. 2. . . . . The MMSE quantizer decision and reconstruction levels are solutions to a set of integral equations. L (2:51) The above MMSE solution. we set the derivatives of D with respect to rk and dk equal to zero. L À 1 with d0 ¼ xmin dL ¼ xmax (2:50) xp(x)dx p(x)dx i ¼ 1.Fundamentals of Digital Image Processing 37 decision levels are determined by minimizing the average distortion (MSE) given by the equation di L X ð i¼1 diÀ1 D¼ (x À ri )2 p(x)dx (2:48) To minimize the average distortion D. However. a closed form solution exists. . In rare occasions (such as the Laplacian distribution). The Lloyd–Max quantizer results in the minimum-mean-square quantization error for a given number of reconstruction levels. . . a numerical solution needs to be determined. . has the following properties: Decision levels are halfway between two adjacent reconstruction levels and reconstruction levels are given by the centroids of the signal probability density that are enclosed between two adjacent decision levels. This yields @D ¼ 2(dk À rk )p(dk ) À 2(dk À rkþ1 )p(dk ) ¼ 0 @dk @D ¼À @rk dk ð (2:49) 2(x À rk )p(x)dx ¼ 0 dkÀ1 Solving these equations results in di ¼ and Ð di ri ¼ Ð iÀ1 di diÀ1 d ri þ riþ1 2 i ¼ 1. Step 2: For the current set of reconstruction levels. also known as the Lloyd–Max quantizer [6].50. compute the optimum decision levels using Equation 2. . The following popular iterative technique due to Lloyd can be used to design an MMSE quantizer with L reconstruction levels: Step 1: Divide the range of the signal values into L uniform reconstruction levels.

SOLUTION The step size for uniform quantizer is Dx ¼ xmax Àxmin ¼ 2 ¼ 0:5. go to Step 2.51. ﬁnd the optimum reconstruction levels by using Equation 2.5.38 Control of Color Imaging Systems: Analysis and Design p(x) 1 –1 1 x FIGURE 2. a. Otherwise.16. For this random variable. Now the process is illustrated using a simple example.5 The PDF of random variable X is shown in Figure 2. If the changes in the reconstruction levels with respect to the previous values are less than a threshold. b. Design the 2 bit (4 levels) Lloyd–Max quantizer and calculate the MSE. Example 2.16 PDF of Example 2. Step 3: For the current set of decision levels. The reconstruction 4 4 and decision levels are d0 ¼ À1 d1 ¼ À0:5 d2 ¼ 0 d3 ¼ 0:5 d4 ¼ 1 r1 ¼ À0:75 r2 ¼ À0:25 r3 ¼ 0:25 r4 ¼ 0:75 The MSE is computed as di L X ð i¼1 diÀ1 À0:5 ð 0 ð D¼ (x À ri ) p(x)dx ¼ À1 2 (x þ 0:75) p(x)dx þ À0:5 1 ð 0:5 ð 2 (x þ 0:25)2 p(x)dx þ 0 (x À 0:25)2 p(x)dx þ 0:5 (x À 0:75)2 p(x)dx or 0:5 ð 1 ð D¼2 0 (x À 0:25)2 (1 À x)dx þ 2 0:5 (x À 0:75)2 (1 À x)dx ¼ 0:0208 . Design the 2 bit (4 levels) uniform quantizer and calculate the MSE. stop.

Fundamentals of Digital Image Processing Since the PDF is symmetric. Therefore. the decision boundaries and the reconstruction levels are d0 ¼ À1 d1 ¼ À0:382 d2 ¼ 0 d3 ¼ 0:382 d4 ¼ 1 r1 ¼ À0:588 r2 ¼ À0:176 r3 ¼ 0:176 r4 ¼ 0:588 . which results in r2 ¼ À0:176. d0 ¼ À1 and d1 ¼ r 1 þ r2 ¼ Àd. the optimum quantizer would be symmetric. that is. 2 d2 ¼ r 2 þ r3 ¼ 0 and 2 d3 ¼ r3 þ r4 ¼d 2 d1 ¼ Àd d2 ¼ 0 d3 ¼ d d4 ¼ 1 39 Therefore the reconstruction levels are r1 ¼ À2d À r2 r3 ¼ Àr2 r4 ¼ 2d þ r2 where r2 and r1 are given by Ð0 r2 ¼ Àd xp(x)dx Ð 0i Àd p(x)dx Ð0 Àd ¼ Ð 0i Àd x(1 þ x)dx (1 þ x)dx ¼ 2d 2 À 3d 6 À 3d and Ð Àd r1 ¼ À1 xp(x)dx Ð Àdi À1 p(x)dx x(1 þ x)dx 3d 2 À 2d 3 À 1 ¼ ÐÀ1 ¼ Àdi 3d 2 À 6d þ 3 À1 (1 þ x)dx Ð Àd The above equations can be reduced to ( r2 ¼ 2d À3d 6À3d 2 À2d À r2 ¼ 3d 2À2d À1 3d À6dþ3 2 3 Eliminating r2 . we have d 4 À 5d 3 þ 8d 2 À 5d þ 1 ¼ 0 Factoring this polynomial yields (d À 1)(d À 1)(d À 0:382)(d À 2:618) ¼ 0 The only acceptable solution is d ¼ 0:382.

2. At each pixel. digital cameras).17. a photodiode converts the incident photons into electrons that are subsequently counted as the pixel intensity. Quantization to 12 or even 14 bits is not . In digital cameras. the sampling is accomplished by a CCD sensor array.g. Example 2.. it is customary to quantize the pixel value to 8 bits (256 levels).. In many imaging systems. 2.g.40 The MMSE is di L X ð i¼1 diÀ1 Control of Color Imaging Systems: Analysis and Design À0:382 ð 0 ð D¼ (x À ri )2 p(x)dx ¼ À1 (x þ 0:588)2 p(x)dx þ À0:382 1 ð 0:382 ð (x þ 0:176)2 p(x)dx þ 0 (x À 0:176)2 p(x)dx þ 0:382 (x À 0:588)2 p(x)dx This can be simpliﬁed as 0:382 ð 1 ð D¼2 0 (x À 0:176) (1 À x)dx þ 2 0:382 2 (x À 0:588)2 (1 À x)dx ¼ 0:01548 Notice that the resulting MSE from Lloyd–Max quantizer is smaller than the MSE obtained in part (a) from uniform quantizer. since it is argued that the human eye cannot distinguish more than 100–200 brightness levels. facsimile) or direct acquisition (e. the following values are obtained for a 3 bit optimum quantizer d0 ¼ À1 d1 ¼ À2:3654 d2 ¼ À1:2474 d3 ¼ À0:5313 d5 ¼ 0:5313 d6 ¼ 1:2474 d7 ¼ 2:3654 d8 ¼ 1 r1 ¼ À3:0658 r2 ¼ À1:664 r3 ¼ À0:83 r4 ¼ À0:2327 r5 ¼ 0:2327 r6 ¼ 0:83 r7 ¼ 1:664 r8 ¼ 3:0658 d4 ¼ 0 The PDF of the Laplacian distribution and its 3 bit quantization levels are shown in Figure 2. pﬃﬃﬃ 2 Àpﬃﬃjxj P(x) ¼ e 2 2 À1 < x < 1 SOLUTION Using the Lloyd–Max iterative algorithm.4 Perceptual Quantization Digital images are obtained either through the process of scanning a raw continuous tone image (e.4.6 Design a 3 bit optimum quantizer for a signal with unit variance Laplacian distribution. The unit variance Laplacian distribution is given by the following PDF.

For example.8]. This approach. let’s assume that each signal value is scalar quantized to 4 bits (16 levels). In the second approach. a small change in the photon count at low intensities would be visible. This suggests an approximate logarithmic relationship between the light intensity and the perceived brightness. to take full advantage of the available quantization levels.Fundamentals of Digital Image Processing P(x) 41 x FIGURE 2.2..7. the pixel value is requantized to 8 bits through a nonlinear digital lookup table (LUT) that simulates a logarithmic or a cube-root transformation. Consequently.g. In general. a high-resolution scanner measures the transmittance of each pixel in roughly 13 bits and uses a cube-root LUT to convert them into 8 bit values. is also less accurate. in many imaging systems the quantization is performed in a perceptual domain as opposed to an intensity domain. while a large change in the photon count at high intensities may not be visible.5 Vector Quantization In VQ. It is well known that uniform quantization of the intensity values (photon counts) usually creates visual artifacts. uncommon in medical imaging systems. There are two methods for implementing the perceptual quantization in imaging systems: In the ﬁrst technique. 2. the analog signal is ﬁrst converted to a perceptual domain through an analog nonlinearity (such as a log ampliﬁer) and the output of the ampliﬁer is quantized to 8 bits. Similarly. the intensity of the pixel value is ﬁrst quantized to a large number of levels (e. the concept of a reconstruction level is generalized to a reconstruction vector and the decision levels are generalized to decision regions. 10–14 bits) during the image digitization stage. Subsequently. VQ introduces less quantization error than SQ for the same number of bits [5. This is because equal changes in the number of photons do not correspond to equal changes in the perceived brightness. although less costly.4. It has been argued that the HVS responds almost to the percentage changes in the intensity. As an example. For example.17 Laplacian distribution with a 3 bit optimum quantizer. which . To achieve the same number of bits. the joint vector should be quantized to 8 bits. some digital cameras initially digitize the pixel intensity to 10 bits and subsequently requantize it to 8 bits using LUTs based on piecewise linear curves simulating a cube-root nonlinearity.

X1 and X2 are uniformly distributed between À1 and þ1. X2 (x1 . x2 ) 6¼ pX1 (x1 )pX2 (x2 ). As can be seen from Figure 2. The reconstruction vectors would be r1 ¼ ½ 0:5 0:5 T and r2 ¼ ½ À0:5 À0:5 T as shown in Figure 2. x2 ) ¼ &1 2 (x1 . The correlation coefﬁcient between x1 and x2 is r¼ 1 E(X1 X2 ) À E(X1 )E(X2 ) 4À0 ¼ qﬃﬃﬃﬃqﬃﬃﬃﬃ ¼ 0:75 sX1 sX2 4 4 12 12 (2:55) If we quantize X1 and X2 separately. x2 ) 2 dashed area otherwise 0 (2:52) The marginal densities of random variables X1 and X2 are 1 ð pX1 (x1 ) ¼ À1 1 ð pX1 . the two quantization vectors c and d are wasted if we use SQ. X2 (x1 .18. In this case. The total average distortion is 0 ð 1 ð D¼ À1 (x1 þ 0:5) p(x1 )dx1 þ (x1 À 0:5)2 p(x1 )dx1 2 0 0 ð 1 ð 1 (x2 þ 0:5)2 p(x2 )dx2 þ (x2 À 0:5)2 p(x2 )dx2 ¼ 6 0 þ À1 (2:56) The quantization levels for X1 and X2 are shown in Figure 2. X2 (x1 .5. To show that VQ outperforms SQ. The design of the MMSE vector quantizer is conceptually similar to the scalar quantizer but is computationally more complex. x2 )dx1 ¼ 0 (2:54) Therefore.5 and 0. The performance of the vector quantizer improves by increasing vector dimension. x2 ) given by pX1 . 2 bits are used and the average distortion is 1. .42 Control of Color Imaging Systems: Analysis and Design provides a choice of 256 reconstruction vectors. The same average distortion can be achieved 6 using 1 bit VQ. the optimum quantizer will be a uniform quantizer with decision boundaries {d1 d2 d3 } ¼ {À1 0 1} and quantization levels of À0. X2 (x1 . x2 )dx2 ¼ &1 2 À1 < x1 < 1 otherwise À1 < x2 < 1 otherwise 0 &1 2 (2:53) pX2 (x2 ) ¼ À1 pX1 . consider two random variables X1 and X2 with a joint PDF pX1 . X2 (x1 . The generalized iterative Lloyd algorithm (also known as the LBG (Linde–Buzo–Gray) or K-means algorithm) is used to design the optimum decision regions and reconstruction vectors. They are not independent since pX1 . each to 1 bit.18.18.

. One approach to quantize X is to quantize each component of vector X separately (SQ). 2.18 VQ and SQ.19 VQ concept when the number of scalars in the vector is 2 and the number of reconstruction vectors is 4. . that is. Quantization Rule: XQ ¼ ri if X is in cell Di i ¼ 1. 2. Let X ¼ [ x1 x2 . 1 i N. An alternative approach is joint quantization of scalars xi ’s. . . . The concept of VQ for N ¼ 2 and L ¼ 4 is shown in Figure 2.Fundamentals of Digital Image Processing x2 1 d –1 1 b r2 –1 c a x1 43 FIGURE 2. L (2:57) where ri . i ¼ 1. The reconstruction vectors ri and the cells Di are determined by minimizing the total distortion deﬁned by mean square quantization error.19. . L are N-dimensional reconstruction vectors and L is the number of levels. Deﬁnition: VQ. . . Â Ã D ¼ E (X À XQ )T (X À XQ ) x2 r4 r3 D3 (2:58) D1 r1 D4 x1 r2 D2 FIGURE 2. . . This is called VQ. . xN ]T be an N-dimensional vector consisting of N scalars xi .

Otherwise. The centroid of Di is P 1 (all vectors in Di ). À2 2 2 3 0 X 3 ¼ 4 2 5. continue to the next step. k X5 À r1k ¼ 3:605. . Centroid: Find a new estimate of the reconstruction vector ri by computing the centroid of the ith cell Di and go to step 2. 2 3 À1 X4 ¼ 4 0 5 2 2 3 2 and X5 ¼ 4 À2 5 0 2 design a 1 bit vector quantizer using K-means algorithm. SOLUTION For a 1 bit quantizer L ¼ 2 and there are two reconstruction vectors. . This is a classiﬁcation problem that can be solved iteratively by K-means algorithm. . They can be picked randomly or we can pick L vectors from the training data set. k X4 À r1k ¼ 2. stop the iteration. . Initialization: Start with an initial estimate of reconstruction vectors ri for i ¼ 1. where Ki is the number of computed by ri ¼ Ki vectors in cell Di .7 Given the ﬁve training vectors: 3 À1 X1 ¼ 4 À2 5. 2. Example 2. or all 1 j L. 2 2 3 À1 X2 ¼ 4 2 5. . . If the average distortion is small (less than a prescribed threshold). we need to ﬁnd the L reconstruction vectors {r1 r2 . i¼1 4. . DL }. k X2 À r1k ¼ 5:656. k X1 À r2k ¼ 5:656 ! X1 2 D1 k X2 À r2k ¼ 0 ! X2 2 D2 k X3 À r2k ¼ 4:123 ! X3 2 D1 k X4 À r2k ¼ 4:472 ! X4 2 D1 k X5 À r2k ¼ 5:385 ! X5 2 D1 . Vector X belongs to Di . . rL } and their corresponding regions {D1 D2 .44 Control of Color Imaging Systems: Analysis and Design K-means Algorithm: Given M training vectors {X1 X2 . Let the initial 2 3 2 3 À1 À1 vectors be r1 ¼ 4 À2 5 and r2 ¼ 4 2 5. . Total distortion: Compute the overall average distortion P 1 D ¼ M M k Xi À XiQ k where XiQ ¼ rk if Xi is in cell Dk . XM }. 2 À2 k X1 À r1k ¼ 0. where k k stands for Euclidean norm. 3. if and only if k X À ri k < k X À rj k . 5. L. 2. . k X3 À r1k ¼ 4:123. Classiﬁcation: Classify the M training vectors into L groups. Following is a summary of the K-means algorithm: 1.

D1 ¼ f X1 The total average distortion is D¼ 1:8708 þ 0 þ 2:549 þ 1:2247 þ 2:915 ¼ 1:7119 5 X3 X4 X5 g D2 ¼ {X2 } k X1 À r2k ¼ 5:656 ! X1 2 D1 k X2 À r2k ¼ 0 ! X2 2 D2 k X3 À r2k ¼ 4:123 ! X3 2 D1 k X4 À r2k ¼ 4:472 ! X4 2 D1 k X5 À r2k ¼ 5:385 ! X5 2 D1 The new reconstruction vectors are 2 3 0 X1 þ X3 þ X4 þ X5 4 ¼ À0:5 5 r1 ¼ 4 1:5 3 À1 r2 ¼ X2 ¼ 4 2 5 À2 2 Carrying out one more iteration does not change the overall average distortion. Image data compression using VQ is a lossy compression technique. k X4 À r1k ¼ 1:2247. Once the codebook has been designed. that are faster than K-means algorithm. D1 ¼ f X1 The total average distortion is D¼ 0 þ 0 þ 4:123 þ 2 þ 3:605 ¼ 1:9456 5 3 À1 r2 ¼ X2 ¼ 4 2 5 À2 2 X3 X4 X5 g D2 ¼ {X2 } 45 The new reconstruction vectors are computed as 2 3 0 X1 þ X3 þ X4 þ X5 4 ¼ À0:5 5 r1 ¼ 4 1:5 Classifying the training vectors with new centroid results in k X1 À r1k ¼ 1:8708. It is also very slow because for every iteration. such as pairwise nearest neighbor and simulated annealing for VQ design. The K-means algorithm is locally optimal and there is no guarantee that it will converge to the global optimal solution. k X5 À r1k ¼ 2:915. the uncompressed image is divided into blocks and each block is converted . Therefore. k X2 À r1k ¼ 4:415. There are other algorithms. k X3 À r1k ¼ 2:549. all the vectors in the data base are compared with each codeword vector.Fundamentals of Digital Image Processing Therefore. An important application of VQ is image data compression.

The most common transforms are discrete Fourier transform (DFT). The decoder will use this index to pick the corresponding codeword from the codebook and generate the output vector. …. …. 2. ri2. …. in much the same way as a continuous . L index i Codebook r1. The encoder will compare each input vector with every codevector in the codebook and output an index corresponding to the minimum distortion codevector from the codebook.5. These vectors are input to the encoder.20 Encoder and decoder for VQ-based image compression technique.5 IMAGE TRANSFORM There are different transforms that can be used in image processing.46 Input image Control of Color Imaging Systems: Analysis and Design x1 x2 x3 x4 Divide into blocks i th block xN Input vector X Encoder Codebook r1. r2. and Karhunen–Loeve (KL) transform [9–12].1 TWO-DIMENSIONAL DISCRETE FOURIER TRANSFORM The 2-D DFT provides an expansion of N Â M samples of a discrete 2-D sequence in terms of discrete sine and cosine signals. discrete cosine transform (DCT). …. riN] ri1 ri2 ri3 ri4 Reconstructed i th block riN FIGURE 2. rL min||X – ri|| i = 1. rL Compressed image Lookup table Output vector ri = [ri1. discrete Hadamard transform. The encoder and decoder for a VQ image data compression system are shown in Figure 2. into a vector. 2. r2. These output vectors are used to reconstruct the compressed image. 2.20.

2. m) 1-D N-Point DFT F (k. Since there are N 2 DFT samples to be computed. l) is N 2 . . .and y-directions. l) FIGURE 2. where fx and fy are the sampling frequencies in the x. the number of complex multiplications needed to compute one sample of F(k. we can compute the 2-D DFT by taking the 1-D DFT of each row of the image and then taking the 1-D DFT of the columns of the intermediate image G(n. . m)eÀj N nk eÀj M ml 2p 2p (2:59) For k ¼ 0. l) ¼ M À1 X X N À1 m¼0 n¼0 f (n.59. m) ¼ X NÀ1 1 MÀ1 X 2p 2p F(k. 1. m) in terms of 2-D complex exponential signals (basis functions). We can compute 2-D DFT more efﬁciently by using the fact that the basis functions involved in DFT expansion are separable. They are generN M ally complex valued and can be used to expand f (n. m) is deﬁned as F(k. The complexity of DFT computation is measured in terms of the number of multiplications required in implementing Equation 2.21 Row–column decomposition. vy ) uniformly spaced at vx ¼ 2p k and vy ¼ 2p l.21. the total number of multiplications is N 4 . N À 1. . The DFT of an N Â M discrete signal f (n. 1-D N-Point DFT G(n. and l ¼ 0. Assuming that N ¼ M. As a result of this. The DFT coefﬁcients are samples of F(vx . . . . as shown in Figure 2. . We refer to this as direct computation. l)ej N nk ej M ml MN l¼0 k¼0 (2:60) k The discrete frequency indices k and l correspond to the analog frequencies N fxs and k s s s N fy . . l). l) f (n. M À 1. respectively. 2. This expansion is called the inverse DFT and is given by f (n. 1.Fundamentals of Digital Image Processing 47 signal can be expanded into sine and cosine using the continuous Fourier transform.

m) . This is due to the fact that the total number of 2 multiplications necessary to compute N-point FFT is N log2 (N). m ¼ r sin u. Conjugation property: If f (n.48 Control of Color Imaging Systems: Analysis and Design TABLE 2. M À l) (2:62) This means that half of the DFT data is redundant data. Now. Therefore. if we use row–column decomposition with fast Fourier transform (FFT) algorithm. u) F(v. the total number of complex multiplications needed is 2N Â N 2 ¼ 2N 3 . m) then 2-D DFT ! F (k. To see the compu2 tational advantage of using FFT for 2-D DFT computation. l) þ bF (k. l) and f (n. k ¼ v cos w. Properties of 2-D DFT The properties of 2-D DFT are listed below. Linearity property: If f1 (n. w) c. a.1 Direct 2-D DFT=FFT Comparison Technique Direct computation Row–column decomposition (1-D DFT) Row–column decomposition (1-D FFT) No. row–column decomposition and row–column decomposition with the FFT algorithm replacing direct DFT computation. m) is a real sequence. l) 1 2 (2:61) b. we need to compute 2N 1-D DFTs and each 1-D DFT requires N 2 multiplications. 2-D DFT 2-D DFT F(k. 2 D DFT 1 2 2 af1 (n. m) þ bf2 (n. With row–column decomposition. then its DFT F(k. of Complex Multiplications N ¼ 7:871947 Â 10 2N 3 ¼ 2:73154048 Â 108 4 10 % 100 0. the total number of multiplications will be reduced to 2N Â N log2 (N) ¼ N 2 log2 (N). Rotation property: If f (n. The proofs are left as an exercise. l) or f (r. m) where ! À ! n ¼ r cos u. l). consider calculating the 2-D DFT of a 512 Â 512 image using direct computation. l) ¼ F*(N À k.1. The comparison is shown in Table 2. m) 2-D DFT ! aF (k.0035 N 2 log2 (N) ¼ 2:359296 Â 106 This is called the row–column decomposition technique.! F (k. l ¼ v sin w .4 0. l) satisﬁes the following conjugation property F(k.

m) and h(n. 0)x(n. ﬁnd 1 a. m) ¼ x(n. m) ¼ XX v u f (u. m). m) h(n. m) ¼ h(0. 1)x(n À 1. m) þ h(1. however. m). m) and h(n. Convolution property of DFT: The convolution theorem for continuous signals states that convolution in space domain is equivalent to multiplication in the Fourier domain. l) by the same angle. m) h(n. l) (2:64) The circular convolution is deﬁned as f (n. This theorem can be extended to the discrete domain. m)*h(n. m À v) (2:65) where the shift (n À u or m À v) is circular shift. 0)x(n À 1. g(n. y(n.m À 1) þ h(1. w þ u0 ) (2:63) This implies that rotating f (n. v)x(n À u mod 2. m) by an angle u0 rotates its transform F(k. b. m). Circular convolution y(n. m) ¼ x(n. m À 1) .1)x(n. m) À ! DFT F(k. f (r. that is. m) ¼ 1 1 XX v¼0 u¼0 h(u. m). circular convolution of x(n. m) ¼ 1 3 ! 1 . linear convolution of x(n. it should be noted that ordinary convolution is replaced by circular convolution. m) ¼ 2 À1 ! 1 1 and h(n. m À 1) þ x(n À 1. u þ u0 ) À ! 2-D DFT F(v. Example 2. The linear discrete convolution is the type of convolution used for ﬁltering signals and images and also for ﬁnding the output of imaging systems modeled in linear shiftinvariant form. m) þ x(n À 1. f (n. m) ¼ x(n. Circular convolution is a side effect of DFT.Fundamentals of Digital Image Processing 49 then. y À v mod 2) Expanding the above equation yields y(n. d.8 If x(n.m) þ h(0. m) h(n. m) þ x(n. m) h(n. SOLUTION a. m À 1) ¼ x(n. v)h(n À u. l)H(k.

À1) ¼ x(1. 1) þ x(0.m) þ x(n. 0) ¼ 3 þ 1 À 1 þ 2 ¼ 5 . 0) ¼ x(1. 1) þ x(À1.m) þ h(0.m À 1) þ h(1. À1) ¼ 2 þ 0 þ 0 þ 0 ¼ 2 y(0. we have y(n. m À 1) Computing y(n. 0) þ x( À1.m À 1) þ x(n À 1. 0) þ x(0. 0) þ x(À1. 0) ¼ 1 þ 2 þ 0 þ 0 ¼ 3 y(0. Linear convolution. 1) þ x(0. À1) þ x( À1. v)x(n À u. y À v ) n ¼ 0. m) for n ¼ 0. 0) þ x(0. 1) þ x(0. 1) ¼ x(1. 1) þ x(0. 2 Expanding the above sum results in y(n. 2. 0) þ x(1. m) þ x(n À 1. 1) þ x(0. 1) þ x(À1. 0) þ x(1. 0) ¼ x(0. 0) þ x(0. 2 m ¼ 0. 1.0)x(n. 1) þ x(0. we would have y(0. 1) þ x(0. 1) þ x(1. 1) þ x( À1. 1) þ x(0. 1) þ x(0. 1. 0) þ x(À1. m) h(n. À1) ¼ x(0. m) ¼ h(0. 1)x(n. À1) þ x(À1. 0) þ x(0. 0)x(n À 1. À 1) þ x(0. 1) ¼ 2 þ 1 À 1 þ 3 ¼ 5 If n ¼ 0 and m ¼ 1. 0) þ x(0. m) ¼ 5 5 b. 0) ¼ 1 þ 2 þ 3 À 1 ¼ 5 If n ¼ 1 and m ¼ 0. 0) þ x(1. 1) þ x(0. 1. 1) ¼ 0 þ 1 þ 0 þ 0 ¼ 1 y(1. 1. À 1) ¼ À1 þ 0 þ 2 þ 0 ¼ 1 y(1. then y(0. then y(0. À1) þ x(0. 2) þ x(À1. then y(1. 1)x(n À 1. 1) ¼ x(0. m) ¼ x(n. m) þ h(1. m À 1) ¼ x(n. 2) ¼ x(0. 0) ¼ x(1. 0) þ x(0. 0) þ x(1. m) ¼ 1 1 XX v¼0 u¼0 ! h(u. 5 5 y(n. 0) ¼ x(0. 0) þ x( À1. m) * h(n. In this case. 1) ¼ x(1. m) ¼ x(n. 0) ¼ 3 þ 1 À 1 þ 2 ¼ 5 Therefore. 1) ¼ x(0. 0) ¼ x(0. 1) þ x(1. then y(1. 2) þ x(0.50 Control of Color Imaging Systems: Analysis and Design If n ¼ 0 and m ¼ 0. 2 and m ¼ 0. 1) ¼ À1 þ 3 þ 2 þ 1 ¼ 5 Finally. 0) þ x(1. if n ¼ 1 and m ¼ 1.

1) þ x(0. Deﬁne N Â N arrays ^(n. m)]} g f d. in general. 0) ¼ x(2. m) ¼ ^(n. 1 f (n. 2) ¼ x(2. m)]DFT[h(n. À1) þ x(1. to perform fast convolution. 0) þ x(1. 0) ¼ 0 þ 0 þ 3 À 1 ¼ 2 y(2.Fundamentals of Digital Image Processing y(1. 1) þ x(1. 1) þ x(2. À1) ¼ 0 þ 0 À 1 þ 0 ¼ À1 y(2. 2) þ x(2. Zero-pad the signal f (n. m) 0 n. 1) þ x(1. Compute ^ ^(n. m). m) ¼ x(n. 0) þ x(2. m M À 1 0 otherwise & h(n. m) is L Â L and we would like to compute g(n. m) ¼ 2 1 ! and SOLUTION a. Choose DFT size to be N ! 3 þ 2 À 1 ¼ 4 and let N ¼ 4. m) ¼ c. m) g MþLÀ1 (2:68) & f (n. Choose N the DFT size to be N ! M þ L À 1 b. 1) ¼ 0 þ 3 þ 0 þ 1 ¼ 4 y(2. m (2:69) Example 2. 2) ¼ x(1. 1) ¼ x(2. one would be interested in computing linear convolution via DFT. the linear convolution of f (n. 1) ¼ 0 þ 0 þ 0 þ 3 ¼ 3 Hence 2 y(n. b. the FFT can be employed for this purpose provided that the data is properly formatted to avoid circular convolution. 2) þ x(0. m) and compute its DFT . m) 0 n. Suppose that signal f (n. As will be shown. a.9 Convolve the following 2-D sequences using DFT. m) ¼ IDFT{DFT[^(n. 2) þ x(1. 0) þ x(1. m) * h(n. m L À 1 0 otherwise (2:66) (2:67) 0 n. m) and h(n. Find g(n. m) is M Â M and the ﬁlter h(n. 2) þ x(1. m) ¼ 4 1 À1 2 3 3 1 5 45 2 3 51 So. m) using DFT. m) ¼ f ^ h(n. m) ¼ 4 2 3 2 3 2 4 3 25 1 4 1 3 h(n.

m) The ith block . . consider designing a convolver for ﬁltering an image f of size M Â M by an ﬁnite impulse response (FIR) ﬁlter of size L Â L.22 Dividing the image into subimages. There are approximately NB ¼ ( M )2 of such image blocks. Zero-pad the signal h(n. l) 2 3 10 12 21 10 7 7 14 14 5 9 4 2 3 9 À39 À j2 7 6 þ j3 j5 7 À27 3 À j6 5 6 À j3 Àj25 1 63 6 g(n. where B ¼ N À L þ 1. m) and compute its DFT 1 62 ^ h(n. In this case...22. l) ¼ 6 4 À j3 7 ! 4 1 À1 À j2 À3 À1 þ j2 5 05 0 4 þ j3 2 À j À2 þ j j5 d. B .. Multiply the 2 DFTs 147 À39 þ j2 6 À22 À j21 j25 ^ ^ ^ G(k. m) ¼ 6 1 f 43 0 2 2 3 1 0 4 2 4 0 Control of Color Imaging Systems: Analysis and Design 3 3 2 21 À5 À j6 9 À5 þ j6 0 6 À1 À j6 À5 À3 1 þ j2 7 0 7 2-D DFT ^ 7 7 ! F(k. As an example. m) ¼ 6 40 0 2 3 1 0 0 0 0 0 0 3 3 2 7 3 À j4 À1 3 þ j4 0 6 Àj5 À2 À j 2 þ j 7 0 7 2-D DFT ^ 7 H(k. l)] ¼ 4 5 6 5 11 17 5 In practice. l)H(k. using FFT hardware capable of performing 1-D N-point FFT. l) ¼ F(k. it may not be practical to compute DFT if the DFT size is large.. l) ¼ 6 4 9 3 þ j6 À22 þ j21 Àj5 ^ e. B M f1 f2 fi .. block convolution can be employed. Take the inverse DFT of G(k. f (n. l) ¼ 6 4 9 À3 9 À3 5 05 À1 þ j6 1 À j2 À3 À5 0 c. m) ¼ IDFT[G(k. M B fi (n. m) FIGURE 2. We ﬁrst divide the image data into blocks of size B Â B..52 1 6 ^(n. as shown in Figure 2.

gi (n.24. l) FIGURE 2. l) N×N – point IDFT N N gi (n. That is. m) ¼ fi (n. m) Zero fi(n. .23 Filtering an image block using DFT. 2. m) using the DFT approach as outlined in Section 2. The output image is formed by overlapping and adding these NB blocks as illustrated in Figure 2. m) N L L h(n .5. m) (2:70) Note that the 2-D DFT computation is performed using row–column decomposition and a 1-D FFT algorithm. .1 and shown in Figure 2. We then convolve each block with h(n.24 Overlap and add technique.23. NB we compute. .Fundamentals of Digital Image Processing N B B fi(n . m) * h(n. B N … … … … FIGURE 2. m) padding 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 53 N N×N – point DFT Fi(k. m) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 N×N – point DFT H(k. . m) Zero padding 0 h (n. for i ¼ 1. .

f6 (n. m) 0 0 3 0 07 7 7 and 05 0 1 61 6 ~ hi (n. f2 (n.54 Control of Color Imaging Systems: Analysis and Design Example 2. m) ¼ 4 15 0 0 0 3 2 9 17 6 19 0 5. divide the image into B2 ¼ 9 blocks 19 f1 (n. we compute ~ gi (n. . m) ¼ 4 14 0 0 0 3 8 0 17 0 5 17 0 3 12 0 10 0 5 18 0 3 16 0 13 0 5 0 0 Next. that is. m) ¼ 4 5 15 7 4 3 2 4 16 0 8 0 5. m) ¼ f (n. where 19 6 5 6 6 12 6 6 10 6 f (n. we convolve each block with h(n. m) ¼ 4 4 9 14 4 3 2 8 8 14 17 17 5. m) ¼ 4 4 12 8 4 3 2 16 18 12 18 1 5. . f5 (n. 2. m) using FFT and inverse FFT. for i ¼ 1. m) ¼ 1 1 1 1 ! SOLUTION First we ﬁnd the block size B B¼NÀLþ1¼4À2þ1¼3 Now. f8 (n. f3 (n.10 Compute g(n. m)]FFT[h(n. 9. . m) ¼ IFFT{FFT[~i (n. m) ¼ 6 6 18 6 6 15 6 4 9 0 2 16 9 12 16 18 15 4 8 19 18 8 18 1 7 16 0 3 4 4 12 5 4 0 15 9 19 9 8 17 11 4 13 3 17 6 8 0 4 17 7 7 14 4 17 7 7 8 14 12 7 7 7 17 6 10 7 7 10 11 18 7 7 14 3 16 5 9 14 13 and h(n. m) ¼ 6 f 6 4 0 fi (n. m) ¼ 4 0 0 2 3 2 16 19 3 9 18 5. m) * h(n. The results are . m)]} f where: 2 6 ~i (n. m) ¼ 4 18 15 2 9 f7 (n. f9 (n. . m) ¼ 4 6 11 10 11 3 2 4 14 3 13 9 5. m) ¼ 6 40 0 2 1 1 0 0 0 0 0 0 3 0 07 7 7 05 0 All FFTs and IFFTs are 4 Â 4 and are computed using row–column decomposition. m) ¼ 4 5 12 2 10 f4 (n. m) using a 4-point FFT.

. 8 7 5 7 3 16 7 16 7 7 7.Fundamentals of Digital Image Processing 2 19 35 49 38 24 26 62 66 30 35 62 47 20 34 53 41 22 19 3 2 3 12 35 36 13 20 42 37 15 4 32 28 0 26 17 3 2 6 14 35 42 21 26 42 45 8 25 34 17 12 22 28 18 16 29 13 0 0 55 3 6 6 24 6 g1 ¼ 6 6 17 4 2 12 10 7 37 7 7 7. 22 9 7 5 0 0 11 29 19 46 27 0 Overlapping and adding blocks g1 through g9 yields 3 19 35 35 19 þ 3 12 26 17 þ 6 14 8 6 24 49 62 37 þ 7 35 45 17 þ 10 35 25 7 7 6 7 6 6 17 38 47 26 þ 8 36 42 14 þ 8 42 34 7 7 6 7 6 6 12 þ 10 24 þ 26 20 þ 34 8 þ 4 þ 18 þ 12 13 þ 20 23 þ 16 14 þ 4 þ 8 þ 14 21 þ 26 17 þ 12 7 7 6 g ¼ 6 28 62 53 19 þ 17 42 50 25 þ 20 42 22 7 7 6 7 6 66 41 8þ9 37 55 27 þ 17 45 28 7 6 33 7 6 6 15 þ 9 30 þ 13 22 þ 20 7 þ 4 þ 16 þ 0 15 þ 4 21 þ 18 10 þ 11 þ 14 þ 3 29 þ 19 18 þ 16 7 7 6 7 6 4 9 21 28 16 þ 15 32 40 23 þ 17 46 29 5 0 8 8 0 þ 15 28 22 9 þ 14 27 13 2 Hence. 42 14 7 5 23 14 16 50 55 21 8 3 6 6 10 6 g3 ¼ 6 6 8 4 4 2 14 6 6 20 6 g6 ¼ 6 6 17 4 3 6 6 17 6 g9 ¼ 6 6 14 4 0 2 7 07 7 7 07 5 0 0 3 6 6 28 6 g4 ¼ 6 6 33 4 9 6 69 6 g7 ¼ 6 60 4 0 2 15 7 19 7 7 7. 19 6 24 6 6 6 17 6 6 6 22 6 g(n. m) ¼ 6 28 6 6 6 33 6 6 24 6 6 4 9 0 2 35 49 38 50 62 66 43 21 8 35 62 47 54 53 41 42 28 8 22 44 34 42 36 17 27 31 15 12 35 36 33 42 37 19 32 28 26 45 42 39 50 55 39 40 22 23 27 22 40 45 44 38 40 14 35 42 47 42 45 48 46 3 8 25 7 7 7 34 7 7 7 29 7 7 22 7 7 7 28 7 7 34 7 7 7 29 5 13 23 27 Example 2. 27 7 5 7 07 7 7 07 5 0 3 0 7 07 7 7 07 5 0 13 20 21 28 8 0 8 0 10 3 18 14 7 40 23 7 7 7. 0 7 5 0 6 6 17 6 g5 ¼ 6 6 9 4 0 6 6 15 6 g8 ¼ 6 6 15 4 0 2 4 7 25 7 7 7. we compare the processing speed of convolution in the spatial domain to that of the same in the frequency domain.11 In this example. 26 7 5 8 18 3 6 67 6 g2 ¼ 6 68 4 2 4 12 7 45 17 7 7 7.

100 FFTs are needed to transform the blocks. (a þ jb)(c þ jd) ¼ ac À bd þ j(ad þ bc) Therefore. algorithm A will be faster than algorithm B if the number of multiplications in A is less than B. 719. the frequency domain convolution is almost twice faster than the spatial domain convolution. The size of 2 each block is (64 À 11 þ 1) Â (64 À 11 þ 1) ¼ 54 Â 54 and the number of blocks is À512Á2 % 100. 424 N1 This implies that. 576 complex multiplications. N2 3 Â 5. 349. For simplicity. 376 ¼ ¼ 0:507 31. 424 ¼ 31:719424 Â 106 Each 2-D FFT requires 64 log2 (642 ) ¼ 24. Since multiplication takes more time than addition. Therefore. 349. . Then. and 100 FFTs to come back to the pixel domain. in this case. Then. 376 complex multiplications Now each complex multiplication is equal to three real multiplications using the following algorithm ac À bd ¼ one multiplication 2 a(c À d) þ one multiplication d(a À b) Two multiplications ad þ bc ¼ d(a À b) þ one multiplication b(d þ c) One additional multiplication Note that the term d(a À b) is computed once. Assume that the 2-D FFT is computed using row–column decomposition. This is equal to a total of 201 FFTs and 409. let’s ignore any symmetrical properties of the ﬁlter as it can be equally taken advantage of in both direct convolution and FFT-based convolution.56 Control of Color Imaging Systems: Analysis and Design Consider a gray scale image of size 512 Â 512 pixels. N2 ¼ 201 Â 24. 600 ¼ 5. Determine which approach will be faster. block convolution using overlap–add method with 64-point FFT or direct convolution in spatial domain? SOLUTION The speed of any digital signal processing algorithm depends on the number of multiplications and additions performed. 600 complex multiplications to multiply the transform coefﬁcients.600 complex multiplications. 576 þ 409. We wish to sharpen the image by processing it with a FIR high-pass ﬁlter of size 11 Â 11. Therefore. Let N1 be the total number of real multiplications when the convolution is performed in spatial domain and N2 be the total number of real multiplications when the overlap–add method is used with a 64-point FFT and inverse FFT. 1 FFT to 54 transform the ﬁlter. N1 ¼ 5122 pixels Â 112 multiplication=pixel ¼ 31. 100 Â 64 Â 64 ¼ 409. 719. Generally speaking. we ignore addition and compute the number of multiplications.

25 4 Â 4 DCT basis functions. . . . The 2-D DCT of 2-D sequence f (n. N À 1 and m ¼ 0. . . . M À 1 The function aN (:) is deﬁned as 8 qﬃﬃﬃ > 1 < N aN (k) ¼ qﬃﬃﬃ > 2 : N k¼0 k 6¼ 0 (2:73) The DCT basis functions for N ¼ 4 are shown in Figure 2.25. 1. l) ¼ aN (k)aM (l) M À1 X X N À1 m¼0 ! ! (2n þ 1)kp (2m þ 1)lp cos f (n. .13]. . m) cos 2N 2M n¼0 (2:71) for k ¼ 0. m) is deﬁned as C(k. . l) cos ! ! (2n þ 1)kp (2m þ 1)lp cos 2N 2M (2:72) for n ¼ 0.Fundamentals of Digital Image Processing 57 2. 1. . 1. u v FIGURE 2. m) ¼ M À1 X X N À1 l¼0 k¼0 aN (k)aM (l)C(k. N À 1 and l ¼ 0. M À 1 The inverse DCT is deﬁned as f (n.5. 1. . .2 TWO-DIMENSIONAL DISCRETE COSINE TRANSFORM The 2-D DCT is used in still image and video data compression [12. . . . . .

l) ¼ 4 0 À1:0031 À1:5 À1:9927 À1:9142 2:8045 2 3 1:3858 À0:9142 7 7 1:4979 5 0:2929 2 15 17 11 12 16 15 12 15 3 14 13 7 7 15 5 13 2.3 TWO-DIMENSIONAL HADAMARD TRANSFORM The 2-D Hadamard transform (HT) is a real transform where the components of the basis functions take values from the binary set {1 À1}.5. m) ¼ 6 4 14 13 The DCT of this image is 56:5 0:574 0 6 3:5042 1:7071 À1:5443 6 C(k.58 Control of Color Imaging Systems: Analysis and Design Example 2.12 Consider the 4 Â 4 subimage 15 6 16 f (n. The N Â N matrix A is called Hadamard transformation matrix and is related to the N Â N Hadamard matrix H by 1 A ¼ pﬃﬃﬃﬃ HN N The 2 Â 2 Hadamard matrix is 1 H2 ¼ 1 1 À1 ! (2:76) (2:75) The 2N Â 2N Hadamard matrix is related to the N Â N Hadamard matrix through the recursive equation given by ! HN HN (2:77) H2N ¼ HN ÀHN For example. The N Â N HT of an N Â N image f is deﬁned as F ¼ AfA (2:74) Here. it is assumed that N ¼ 2m . ! 1 61 ¼6 41 1 2 1 1 À1 1 1 À1 À1 À1 3 1 À1 7 7 À1 5 1 H4 ¼ H2 H2 H2 ÀH2 (2:78) .

which implies that AÀ1 ¼ A. and 2 1 1 1 1 À1 À1 1 1 À1 À1 1 1 1 1 1 À1 À1 À1 À1 1 À1 1 À1 À1 1 À1 1 1 1 3 H8 ¼ H4 H4 6 1 À1 1 6 6 6 1 1 À1 ! 6 6 1 À1 À1 H4 6 ¼6 61 1 1 ÀH4 6 6 1 À1 1 6 6 4 1 1 À1 1 À1 À1 1 7 7 7 À1 À1 7 7 À1 1 7 7 7 À1 À1 7 7 À1 1 7 7 7 1 1 5 À1 1 À1 (2:79) The basis functions for the 4 Â 4 HT are shown in Figure 2.1 Inverse Hadamard Transform The columns of matrix A are forming a set of orthogonal vectors of unit norm. 2.26 4 Â 4 HT basis functions. This implies that A is unitary.5.Fundamentals of Digital Image Processing 59 FIGURE 2.3.26. pre-multiply and post-multiply Equation 2. Now to ﬁnd the inverse HT.74 by AÀ1 AÀ1 FAÀ1 ¼ f f ¼ AFA (2:80) (2:81) .

and image deblurring.60 Control of Color Imaging Systems: Analysis and Design Example 2. Some examples of popular small size FIR ﬁlters used in image processing are a. a kernel. m)A 2 1 1 1 6 1 À1 ¼ 6 241 1 1 À1 2 32 15 15 1 1 1 À1 76 16 17 76 À1 À1 54 14 11 13 12 À1 1 3 0 0 À1 2 7 7 2:5 À2:5 5 À1:5 À1:5 16 15 12 15 3 2 14 1 1 1 13 7 1 6 1 À1 1 7 6 15 5 2 4 1 1 À1 13 1 À1 À1 3 1 À1 7 7 À1 5 1 56:5 1:5 6 À0:5 À0:5 F(k. or elements. no multiplications are needed in the computation of the HT. Since the basis vectors contain only Æ1 values. weights. m) given as 15 6 16 f (n. l) ¼ Af (n.13 Find the HT of the 4 Â 4 image f (n. image sharpening.6 IMAGE FILTERING Filtering images has applications in image smoothing (noise removal). m) ¼ 6 4 14 13 2 15 17 11 12 16 15 12 15 3 14 13 7 7 15 5 13 SOLUTION F(k. l) ¼ 6 4 4 0 0 1 The HT is a real and fast transform. edge enhancement. m) ¼ 6 1 94 1 2 1 1 1 1 1 2 1 1 3 h(n. or a window [14]. 5 1 1 1 6 h(n. The ﬁlter components are referred to as ﬁlter taps. Filters used in image processing are generally FIR ﬁlters of relatively small size. 5 1 1 3 1 6 62 16 4 1 1 2 4 2 1 3 7 2 7. An FIR ﬁlter is also referred to as a mask. 2. m) ¼ 2 7 1 7. Low-pass ﬁlters (LPF) 2 16 h(n. m) ¼ 6 1 10 4 1 7 17 5 1 (2:82) .

Fundamentals of Digital Image Processing 61 b. These types of ﬁlters are referred to as LPF. and the ﬂatter the weights. The larger the ﬁlter. 2) in the following 8 Â 8 block of the LENA image 139 6 144 6 6 6 150 6 6 159 f (n. Examples of ﬁlters are LPF: If the ﬁlter weights are all positive. noise removal. m) ¼ 6 À1 5 4 0 À1 2 À1 À2 16 h(n. and the results are summed together. 5 1 À2 2 À1 À1 6 h(n. m) ¼ 6 À2 4 1 À2 5 7 À2 7. 2) ¼ 1 (4 Â 161 þ 2 Â 161 þ 2 Â 161 þ 2 Â 161 þ 2 Â 161 þ 1 Â 162 16 þ 1 Â 162 þ 1 Â 160 þ 1 Â 163) ¼ 161 (2:86) 2 4 2 3 1 25 1 (2:85) FIR ﬁlters are used in many image processing tasks such as smoothing. the smoother the resulting image. 2) is computed as g(5. the ﬁltering operation creates a smoother (softer) image that consists mostly of low frequencies. m) ¼ 6 À2 19 74 À1 À2 7 À1 7. High-pass ﬁlters (HPF) 2 0 À1 6 h(n. m) ¼ 16 1 the output pixel g(5. and edge detection [15–17]. ------------- ------------- . 5 0 À1 À1 3 0 3 2 6 h(n. Each ﬁlter tap is multiplied by its corresponding pixel value. LPF are used as anti-aliasing ﬁlters or prior to subsampling in image decimation. As an example. consider ﬁltering the pixel f (5. m) ¼ 6 À1 9 4 À1 À1 7 À1 7 5 The FIR ﬁltering of an image is performed one pixel at a time. 5 1 À1 À1 3 (2:83) 1 3 7 À2 7. m) ¼ 6 6 159 6 6 6 161 6 4 162 162 2 144 149 153 151 153 156 155 160 163 161 162 160 --------------------160 161 162 161 161 161 162 161 163 --------------------162 161 161 155 159 158 159 162 160 162 163 155 156 156 159 155 157 157 158 155 156 156 158 155 157 157 158 3 155 156 7 7 7 156 7 7 159 7 7 155 7 7 7 157 7 7 157 5 158 (2:84) Using the 3 Â 3 FIR ﬁlter 2 1 1 4 2 h(n. The center ﬁlter weight is placed on the pixel to be ﬁltered. sharpening.

LPF 2 6 62 6 1 6 h(n. so as to maintain the average brightness of the output image at the same level as the input image. HPF 2 7 6 6 À2 þ4 À8 þ4 À2 7 7 6 7 6 h(n. 7 7 2 27 5 2 1 1 1 3 6 61 6 1 6 h(n. m) ¼ 6 2 60 6 6 62 4 1 6 61 6 1 6 h(n. m) ¼ 6 þ4 À8 13 À8 þ4 7 7 6 7 6 6 À2 þ4 À8 þ4 À2 7 5 4 þ1 À2 þ4 À2 þ1 3 2 À1 À1 À1 À1 À1 7 6 6 À1 À1 À1 À1 À1 7 7 6 7 16 h(n. the ﬁlter taps in both LPF and HPF sum up to 1. m) ¼ 6 À1 À1 28 À1 À1 7 7 6 46 7 6 À1 À1 À1 À1 À1 7 5 4 À1 À1 À1 À1 À1 þ1 À2 þ4 À2 þ1 3 2 1 1 2 2 4 2 2 1 2 3 2 1 2 4 8 4 2 2 3 5 3 2 2 1 3 2 7 2 27 7 7 4 2 7. These ﬁlters are referred to as HPF because they generally boost the high-frequency components of the image. The larger the relative contribution of the negative taps of the ﬁlter. m) ¼ 6 2 45 6 6 61 4 1 b. m) ¼ 6 1 25 6 6 61 4 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 3 7 17 7 7 1 7. Following ﬁlters are examples of common FIR ﬁlters used in image processing: a. the sharper is the ﬁltered image. and some or all of the ﬁlter weights around the center weight are negative.62 Control of Color Imaging Systems: Analysis and Design HPF: The weight at the center is positive. This kind of ﬁlter often creates a sharper image by accentuating the differences among pixels. Usually. 7 7 17 5 1 7 2 17 7 7 3 27 7 7 2 17 5 1 1 .

quantize the magnitude of the gradient image estimated by . Filter to detect vertical line À1 h(n. consider edge detection based on the two ﬁlters speciﬁed in g and h above. Filter to detect lines oriented at À45 2 2 À1 h(n. Filter to detect horizontal line À1 À1 h(n. respectively. m) ¼ 4 0 1 2 h. Filter for detecting vertical edges À1 hv (n. That is. m) ¼ 4 À1 À1 2 2 2 2 3 À1 À1 5 À1 (2:88) 2 3 À1 2 5 À1 (2:87) e. Filter to detect lines oriented at 45 2 À1 À1 h(n. m) ¼ f (n. Filter for detecting horizontal edges 2 À1 À2 0 hh (n.and y-directions by ﬁltering the image using ﬁlters hh (n. m) ¼ 4 2 2 À1 À1 d. m) ¼ 4 À1 2 2 À1 f. m) * hv (n. m) ¼ 4 À1 2 À1 À1 g. To create an edge map of image f .Fundamentals of Digital Image Processing 63 c. m) ¼ 4 À2 À1 2 0 0 0 3 2 À1 5 À1 (2:89) 3 À1 À1 5 2 (2:90) 3 À1 0 5 1 (2:91) 3 1 25 1 (2:92) As an example of FIR ﬁltering. m) * hh (n. m) ¼ f (n. m) (2:93) (2:94) To ﬁnd the edge map of the image. m) and hv (n. m) Gy (n. These ﬁlters are known as Sobel operators. we ﬁrst compute the gradient of f in the x. m). Gx (n.

Àn2 ) (2:97) The ﬁrst step in designing 2-D FIR ﬁlters using the window method is to compute the desired ﬁlter impulse response by inverting the desired frequency response using the inverse Fourier transform. that is. n2 ) ¼ 1 4p2 p p ð ð Hd (jv1 . m) T (2:95) G(n. m) Control of Color Imaging Systems: Analysis and Design EMAP (n. that is. 2.1 DESIGN OF 2-D FIR FILTERS Two-dimensional FIR ﬁlters can be designed to have zero-phase response.6. Zerophase ﬁlters introduce no phase distortion. m)j þ jGy (n. m) ¼ jGx (n. . m)j to 1 bit using the threshold T & EMAP(n. hd (n1 . G(n. The necessary and sufﬁcient condition for 2-D FIR ﬁlters to have zero-phase response is that the ﬁlter impulse response be symmetric with respect to the origin (center point). n2 ) ¼ h(Àn1 . HP. m) > T (2:96) The result of applying the above edge detection algorithm to the LENA image is shown in Figure 2.4 (a) (b) FIGURE 2.64 f (n. and band-pass ﬁlters for both rotationally symmetrical and nonsymmetrical ﬁlters [5].27b. jv2 )ej(v1 nþv2 n2 ) dv1 dv2 (2:98) Àp Àp Closed form solutions exist for the 2-D impulse response sequences of LP.27 (a) Original LENA image and (b) edge map of LENA image. m) T = 0. h(n1 . m) ¼ 1 0 G(n.

for symmetrical ﬁlters the 2-D window is qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ n2 þ n2 w(n1 . n2 ) À 2p n2 þ n2 1 2 Band pass (circularly symmetric) pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2 2 2 2 vc2 J1 vc2 n1 þ n2 vc1 J1 vc1 n1 þ n2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ À 2p 2p n2 þ n2 n2 þ n2 1 2 1 2 They are tabulated in Table 2. For example. n2 ) (2:99) The 2-D window w(n1 . n2 ) À sin (vcx pn1 ) sin (vcy pn2 ) pn1 pn2 Band pass sin (vcx2 pn1 ) sin (vcy2 pn2 ) pn1 pn2 sin (vcx1 pn1 ) sin (vcy1 pn2 ) À pn1 pn2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2 2 vc J1 vc n1 þ n2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2p n2 þ n2 1 2 Low pass (circularly symmetric) ωc High pass (circularly symmetric) pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 2 2 vc J1 vc n1 þ n2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ d(n1 . which is h(n1 . n2 ) is the entire (n1 . it has to be truncated for a ﬁnite support. n2 ) plane.2 Ideal Impulse Response of Symmetric and Nonsymmetric Filters Filter Type Low pass hd (n1 . The truncation is done using a 2-D window function. n2 )w(n1 . n2 ) sin (vcx pn1 ) sin (vcy pn2 ) pn1 pn2 Support of Hd ( jv1 . The support for the impulse response hd (n1 . n2 ) ¼ hd (n1 .Fundamentals of Digital Image Processing 65 TABLE 2. jv2 ) High pass d(n1 . n2 ) is generally designed using 1-D windows.2. n2 ) ¼ w1 1 2 (2:100) . Therefore.

29. This is due to the side lobes in the spectrum of the uniform window. The 1-D uniform. The purpose of a window is to smooth the frequency response. v2 ) (2:102) This means that the frequency response of the designed ﬁlter is the convolution of the desired frequency response and the frequency response of the truncated window. Hanning. The most frequently used 1-D windows and their spectral properties are listed in Table 2.28 and 2.14 Design a circularly symmetric 2-D low-pass FIR ﬁlter for an image of size 57 Â 57 with a cutoff frequency of vc ¼ p. To minimize the effect of these side lobes and reduce the ripples in the passband. . Hamming. The frequency response of the designed ﬁlter can be obtained by taking the Fourier transform from both sides of Equation 2. v2 ) ¼ Hd (v1 . we can use other windows such as Hamming.66 Control of Color Imaging Systems: Analysis and Design TABLE 2. 4 . .3. Example 2. spatial and frequency domains are shown in Figures 2. Using the convolution property of the Fourier transform. the designed ﬁlter will have ripples in its passband.99. we get H(v1 . respectively. The window must have a small mainlobe width so that the transition width of H(v1 . À M þ 1. or Blackman windows. . M 1 np 0:54 þ 0:46 cos M np 0:5 þ 0:5 cos M np 2np þ 0:08 cos 0:42 þ 0:5 cos M M Distance between the Peak of the Mainlobe and the Peak of the First Sidelobe (dB) 13 43 46 50 For separable ﬁlters. v2 ) is small. n2 ) ¼ w1 (n1 )w2 (n2 ) (2:101) The functions w1 (n1 ) and w2 (n2 ) are 1-D windows.3 Spectral Property of Different Windows Window Type Uniform Hamming Hanning Blackman w(n). If the truncated window is a uniform window. We also need to have a window with the smallest sidelobe amplitude to make the ripple in the passband and stop-band regions as small as possible. n ¼ ÀM. v2 ) * W(v1 . w(n1 . . Hanning. and Blackman windows of size 32 in.

5 ω 2 2.8 w(n) Hanning 0.5 3 3.5 1 1. n2 ) ¼ ¼ 2p 8 n2 þ n2 n2 þ n2 2 2 1 1 where J1 ( Á ) is Bessel function of the ﬁrst kind and ﬁrst order.28 Windows in spatial domain.4 0 0 5 10 15 n 20 25 30 FIGURE 2.29 Windows in frequency domain.6 Hamming Blackman 0. 0 Uniform –50 –100 –150 W(jω)| –200 –250 –300 –350 –400 Blackman Hanning Hamming 0 0. The designed ﬁlter using a hamming window would be .2 0.Fundamentals of Digital Image Processing 67 1 Uniform 0.5 FIGURE 2. SOLUTION The desired impulse response is given by pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ p 2 2 2 2 vc J1 vc n1 þ n2 1 J1 4 n1 þ n2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ hd (n1 .

5 –1 –1 –0. À27. 28 The frequency responses of the designed ﬁlter using uniform and hamming windows are shown in Figure 2.5 0.9 1. where the two rates are related by a factor of a. n2 ) ¼ 2 þ n2 8 28 n1 2 n1 .5 0. . pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ " pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ !# p 2 2 1 J1 4 n1 þ n2 n2 þ n2 p 2 1 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 0:54 þ cos h(n1 . R2 ¼ aR1 (2:103) . . 0. .2 0.1 FIGURE 2. 2. n2 )w(n1 . Sampling rate conversion decouples the spatial resolution of the image source from the spatial resolution requirement of the display or printing device.5 0 Fy –0. that is.30. n2 ) ¼ hd (n1 . Similarly. for the purpose of viewing this image on a high deﬁnition television (HDTV) monitor.7 IMAGE RESIZING In many imaging applications. This image is suitable for high quality printing.6 0.68 Control of Color Imaging Systems: Analysis and Design 1 0. . Image resizing is basically performed by the process of sampling rate conversion.3 0.8 0. . An example for such a situation is when an image is initially scanned at a high resolution (for example: 3072 Â 2048 pixels). or printing. create a new image at a spatial rate R2 . 2. 1. Thus.5 0.4 0 1 0.5 0. an even lower resolution image is needed for national television system committe (NTSC) television viewing.30 Frequency response of the design ﬁlter. a lower resolution image is required. However. the same image source can be viewed on different display systems with different resolutions [18.7 Magnitude 1 0.1 DEFINITION OF SAMPLING RATE CONVERSION Given an image sampled at a rate R1 .7.19]. n2 ¼ À28. . . transmitting.5 0 Fx 1 0. it is sometimes necessary to create a low-resolution image from a high-resolution image and vice versa for purposes of viewing. .

we ﬁrst upsample (interpolate) the image by a factor of P and then downsample (decimate) by a factor of Q. Extension to 2-D signals is straightforward. The output signal before low-pass ﬁltering is related to the input by & ÀnÁ f P y(n) ¼ 0 The DTFT of y(n) is given by Y(v) ¼ 1 X n¼À1 n ¼ 0. otherwise (2:105) y(n)eÀjnv ¼ 1 1 X n X f f (k)eÀjkPv ¼ F(Pv) eÀjnv ¼ P n¼À1 À1 (2:106) Assume that f (n) is obtained by sampling a truly band-limited analog signal fa (x) at 1 the Nyquist rate of fs ¼ Dx. Here. 2. then the discrete time Fourier transform of g(n) will be y(n) f (n) P H( jω) g(n) FIGURE 2.2 UPSAMPLING BY FACTOR OF P To upsample a 1-D signal by a factor of P. the sampling rate can be changed by a factor of a only if it is a rational number (i. The second approach is digital processing. We ﬁrst consider the upsampling process.7. that is 1 1 X v À 2pk Fa j F(jv) ¼ Dx k¼À1 Dx (2:107) P If we now sample fa (x) at the rate of Pfs ¼ Dx. this process will be presented for 1-D signals.e. . we insert P À 1 zeros between the signal samples and LPF the resulting signal.. the continuous image is ﬁrst reconstructed and then it is resampled at the new desired rate. The ﬁrst technique is continuous domain processing. ÆP.31 Upsampling by a factor of P. . In discrete domain. . which implies that the image is processed in the digital domain without conversion to analog or continuous domain. . that is a¼ P Q (2:104) In this case.Fundamentals of Digital Image Processing 69 There exist two approaches to sampling rate conversion. ratio of two integers). Æ2P. This approach is theoretically possible but is not practical. to obtain g(n). In this approach. The block diagram of an upsampler is shown in Figure 2.31.

70

**Control of Color Imaging Systems: Analysis and Design
**

H( jω)

P

ω –π P π P

**FIGURE 2.32 Frequency response of ideal interpolation ﬁlter.
**

1 P X v À 2pk G(jv) ¼ Fa j Dx Dx k¼À1 P

! (2:108)

Therefore, if the signal y(n) is passed through an ideal LPF with a gain of P and cutoff frequency of p, the output would be the desired upsampled signal g(n). This P means that the ideal interpolation LPF has the frequency response given by & H(jv) ¼ P 0 jvj < p P (2:109)

otherwise

This frequency response is shown in Figure 2.32.

2.7.3 DOWNSAMPLING

BY

FACTOR

OF

Q

To downsample a 1-D signal by a factor of Q, we retain every Qth sample. The block diagram of a downsampler is shown in Figure 2.33. The output signal is related to the input by y(n) ¼ f (Qn) The DTFT of y(n) is given by Y(v) ¼

1 X n¼À1

(2:110)

y(n)eÀjnv ¼

1 X n¼À1

f (Qn)eÀjnv ¼

1 X k¼À1

f (k)w(k)eÀjQv

k

(2:111)

f (n)

Q

y(n)

FIGURE 2.33 Downsampling by a factor of Q.

Fundamentals of Digital Image Processing

71

where w(k) is deﬁned as & w(k) ¼ 1 0

1 X k ¼ 0, ÆQ, Æ2Q, . . . d(k À nQ) ¼ otherwise n¼À1

(2:112)

**Since w(k) is periodic with a period of Q, it can be expanded using discrete Fourier series (DFS) as w(k) ¼
**

QÀ1 1 X j2pnk eQ Q n¼0

(2:113)

**Substituting Equation 2.113 into 2.111 yields Y(v) ¼
**

1 X k¼À1

f (k)w(k)eÀjQv ¼

k

1 X k¼À1

f (k)

QÀ1 1 X v À 2pn ¼ F Q n¼0 Q

QÀ1 QÀ1 1 k vÀ2pn 1 X j2pnk ÀjQv 1 X X e Q e ¼ f (k)eÀj Q k Q n¼0 Q n¼0 k¼À1

(2:114)

As evident from Equation 2.114, there are Q terms in the expansion of Y(v). This p expansion shows that there will be aliasing if the signal bandwidth is more than Q. To avoid aliasing, we need to LPF the signal before downsampling. The frequency response of the ideal LPF is & H(jv) ¼ 1 0

p jvj < Q otherwise

(2:115)

The frequency response of the ideal anti-aliasing LPF is shown in Figure 2.34. The overall block diagram of a downsampler is shown in Figure 2.35.

H( jω)

1

ω π – Q π Q

FIGURE 2.34 Frequency response of anti-aliasing ﬁlter.

72

**Control of Color Imaging Systems: Analysis and Design
**

y(n) f (n) H( jω) Q

g (n)

FIGURE 2.35 Downsampling by a factor of Q.

2.7.4 SAMPLING RATE CONVERSION

BY A

FACTOR

P OF Q

P To change the sampling rate by a factor of Q, we ﬁrst upsample the signal by a factor of P and then downsample it by a factor of Q, as shown in Figure 2.36. The two LPFs can be combined into one ﬁlter with a passband gain of P Â 1 ¼ P p and a bandwidth given by the minimum of p and Q. The overall system is shown in P Figure 2.37. In this ﬁgure, the LPF bandwidth is given by

& BW ¼ min

p p , P Q

' (2:116)

2.7.5 EXAMPLES

OF

LOW-PASS FILTERS USED

FOR

SAMPLING RATE CONVERSION

An ideal LPF is not realizable, so in practice it is approximated by a simple realizable FIR ﬁlter. For example, the following simple ﬁlters can be used for upsampling. These ﬁlters are derived using a linear interpolation algorithm. For example, for upsampling by a factor of P, the LPF would be an FIR ﬁlter of size 2P À 1, given by 1 h(n) ¼ P 2 P ÁÁÁ PÀ1 P 1 PÀ1 P ÁÁÁ 2 P ! 1 P (2:117)

**The DC gain of this ﬁlter is DC gain ¼ H(jv)jv¼0 ¼ X
**

n

h(n) exp (Àjv)jv¼0 ¼

X

n

h(n) ¼ P

(2:118)

f (n)

P

H1(jω)

H2(jω)

Q

g (n)

FIGURE 2.36

P Sampling rate conversion by a factor of Q.

H(ω) P f (n) P –BW BW Q ω g (n)

P FIGURE 2.37 Sampling rate conversion by factor of a Q.

Fundamentals of Digital Image Processing

73

2 Ideal

1.5 |H(jω)| Linear interpolation 1

0.5

0 0 0.5 1 1.5 ω 2 2.5 3

FIGURE 2.38 Frequency response of ideal and linear interpolation ﬁlters for P ¼ 2.

**For P ¼ 2 and P ¼ 3, the ﬁlters are h(n) ¼ Â1
**

2

1

1 2

Ã

and

h(n) ¼

Â1

3

2 3

1

2 3

1 3

Ã

(2:119)

The frequency response of the interpolation ﬁlter for P ¼ 2 is shown in Figure 2.38. For downsampling by a factor of Q, the same ﬁlter can be used except that the DC gain of the ﬁlter should be one. Therefore, 1 1 h(n) ¼ Q Q 2 Q ÁÁÁ QÀ1 Q 1 QÀ1 Q ÁÁÁ 2 Q 1 Q ! (2:120)

Example 2.15

Given a 4 Â 4 sequence 2 64 6 f (n, m) ¼ 4 2 2 2 3 1 4 1 1 3 1 3 3 2 57 7 25 2

upsample the sequence f (n, m) by a factor of 2 using linear interpolation.

SOLUTION

We ﬁrst insert zeros between pixels and increase the size to 8 Â 8.

74

Control of Color Imaging Systems: Analysis and Design 2 60 6 6 64 6 60 g(n, m) ¼ 6 62 6 6 60 6 42 0 2 0 0 0 0 0 0 0 0 3 0 1 0 4 0 1 0 0 0 0 0 0 0 0 0 1 0 3 0 1 0 3 0 0 0 0 0 0 0 0 0 2 0 5 0 2 0 2 0 3 0 07 7 7 07 7 07 7 07 7 7 07 7 05 0

Now we ﬁlter the signal using a separable 2-D Ã ﬁlter. The row and column ﬁlters Â are identical and each is given by h ¼ 1 1 1 . We ﬁrst ﬁlter each row using h. 2 2 The resulting image is 2 2:5 60 0 6 6 6 4 2:50 6 60 0 g2 (n, m) ¼ g1 (n, m)*h(m) ¼ 6 62 3 6 6 0 60 6 4 2 1:5 0 0 2 3 2 1 1:50 2 0 0 0 0 0 1 2 3 4 5 0 0 0 0 0 4 2:5 1 1:5 2 0 0 0 0 0 1 2 3 2:5 2 0 0 0 0 0 3 1 0 7 7 7 2:5 7 7 0 7 7 1 7 7 7 0 7 7 1 5 0

Apply the same ﬁlter h across each column of image g2 (n, m) and this process results in 2 2:5 3 2 6 3 2:5 2 2 6 6 2 6 4 2:50 1 6 6 3 2:75 2:5 2:25 g(n, m) ¼ g2 (n, m) * h(n) ¼ 6 62 3 4 2:5 6 6 6 2 2:25 2:5 2:25 6 4 2 1:5 1 2 1 0:75 0:5 1 2 3 1 1:50 2 1 2 2:75 3:5 1:75 7 7 7 3 4 5 2:5 7 7 2 2:75 3:5 1:75 7 7 1 1:5 2 1 7 7 7 2 2 2 1 7 7 3 2:5 2 1 5 1:5 1:25 1 0:5

Example 2.16

Resize the following 6 Â 6 image f (n, m) to a 4 Â 4 image. 1 63 6 62 f (n, m) ¼ 6 65 6 42 4 2 2 0 3 3 0 3 4 2 1 2 0 2 2 4 0 1 5 7 1 3 6 0 6 3 3 2 27 7 47 7 37 7 25 2

Fundamentals of Digital Image Processing

75

SOLUTION

We need to change the sampling rate by a factor of 2. For this, we ﬁrst upsample 3 the image by a factor of 2 and then downsample it by a factor of 3. In both cases, we use the linear interpolation ﬁlter given by Equation 2.117. The result of upsampling by a factor of 2 is 2 1 2 3 1:5 1:5 1:5 2 1 0 3 2 1 4 3 2 3 3 3 2 3 4 1:5 2:5 3:5 1 2 3 1:5 2 2:5 2 2 2 1 1 1 3

6 6 6 6 6 6 2:5 2 1:5 1:5 1:5 1:75 2 3:25 4:5 6 6 6 2 2:5 3 2 1 0:5 0 3 6 6 6 3:5 3:25 3 2:25 1:5 1 0:5 1:75 3 6 g1 (n, m) ¼ 6 6 5 4 3 2:5 2 1:5 1 0:5 0 6 6 3:5 2:5 1:5 1:25 1 2 3 3 3 6 6 6 2 1 0 0 0 2:5 5 5:5 6 6 6 3 2:25 1:5 1:25 1 3:5 6 5:25 4:5 6 6 4 4 3:5 3 2:5 2 4:5 7 5 3 2 1:75 1:5 1:25 1 2:25 3:5 2:5 1:5

7 7 7 7 7 3:75 3 1:5 7 7 7 5 4 2 7 7 3:25 3:5 1:75 7 7 7 1:5 3 1:5 7 7 2:75 2:5 1:25 7 7 7 4 2 1 7 7 3:25 2 1 7 7 7 2:5 2 1 5 1:25 1 0:5

Now before downsampling g1 (n, m) by a factor of 3, we LPF the signal using a Â Ã separable ﬁlter constructed from the 1-D ﬁlter h(n) ¼ 1 2 1 2 1 . Applying 9 9 3 9 9 this ﬁlter across rows and columns of g1 (n, m) will result in

0:69 6 6 1:02 6 6 1:28 6 6 6 1:49 6 6 1:8 6 6 6 2:10 g2 (n, m) ¼ 6 6 2:15 6 6 6 1:90 6 6 1:68 6 6 6 1:64 6 6 1:56 4 1:14 2 0:95 1:23 1:54 1:83 1:89 1:73 1:48 1:31 1:19 0:99 0:65 1:27 1:49 1:77 2:17 2:41 2:43 2:29 2:13 1:9 1:52 1:63 1:76 2:07 2:41 2:69 2:88 2:93 2:67 1:80 1:83 1:71 1:72 1:92 2:35 2:92 3:34 3:22 2:24 2:26 1:9 1:57 1:49 1:83 2:54 3:21 3:32 1:6 1:39 1:55 2:09 2:71 2:94 2:58 2:56 2:07 3

2:54 2:41 1:93 1:63 1:61 1:81 2:17 2:52 2:61 2:14 1:93 1:57 1:64 2:10 2:64 2:97 3:01 2:71 1:85 1:63 1:41 1:80 2:72 3:62 3:98 3:73 3:02 1:87 1:71 1:57 2:09 3:15 4:10 4:34 3:85 2:94 1:84 1:77 1:66 2:10 2:98 3:69 3:70 3:12 2:30 1:38 1:36 1:28 1:54 2:07 2:45 2:35 1:90 1:37

7 1:51 0:95 7 7 2:07 1:27 7 7 7 2:54 1:53 7 7 2:74 1:70 7 7 7 2:56 1:66 7 7 2:26 1:48 7 7 7 2:12 1:31 7 7 2:12 1:21 7 7 7 1:97 1:09 7 7 1:53 0:86 7 5 0:93 0:54

Now we downsample the signal g2 (n, m) by a factor of 3. 3 0:69 1:54 1:73 1:19 6 1:49 1:71 2:35 3:22 7 7 6 g(n, m) ¼ 6 7 4 2:15 1:93 1:81 2:61 5 1:64 1:57 4:10 2:94 2

76

Control of Color Imaging Systems: Analysis and Design

FIGURE 2.39 256 Â 256 LENA image.

FIGURE 2.40 Upsampled LENA image.

Example 2.17

In this example, the LENA image is upsampled by a factor of 2 and downsampled by a factor 3. The linear interpolation ﬁlters are used for both the upsampling and downsampling operations. The original image is shown in Figure 2.39. The upsampled and the downsampled images are shown in Figures 2.40 and 2.41, respectively.

Fundamentals of Digital Image Processing

77

FIGURE 2.41 LENA image at a resolution of 2 of the original image. 3

2.8 IMAGE ENHANCEMENT

The goal of image enhancement is to provide a more instinctively pleasing image. Enhancements are also used to simplify visual interpretation and understanding of images. There are different techniques used in image enhancement such as unsharp masking and histogram equalization [5,19–21].

2.8.1 UNSHARP MASKING

The operation of unsharp masking is used to sharpen an image by enhancing its highfrequency components. The process consists of the following three steps. 1. Image f (n, m) is slightly blurred to create a smooth (low-pass) image (n, m). f 2. Difference between the original image and the smooth image is d(n, m) ¼ f (n, m) À (n, m). This image consists mainly of high frequencies of the f original image. 3. Difference image is boosted by a factor of b and added back to the original image to create a sharper image. The degree of sharpness is controlled by the boost factor and the size of the LPF f fUM (n, m) ¼ f (n, m) þ b(f (n, m) À (n, m)) (2:121)

The block diagram representing the unsharp masking process is shown in Figure 2.42. In digital processing, the entire process of linear unsharp masking can be compactly represented as a single high-pass FIR ﬁlter operation. Consider performing unsharp masking using an FIR ﬁlter hLP (n, m) as the blurring operator and a boost factor of b. The ﬁlter used in this process is denoted by d(n, m) and has all components as zeros except for the center tap (which is one),

78

**Control of Color Imaging Systems: Analysis and Design
**

– – f (n, m) d(n, m) + fUM(n, m) +

f (n, m)

Low-pass filter

β

+

FIGURE 2.42 Unsharp masking.

the equivalent FIR ﬁlter starting from the original image f (n, m) to the output unsharp masked image fUM (n, m) is given by hUM (n, m) ¼ (1 þ b)d(n, m) À bhLP (n, m) For example, if the LPF hLP (n, m) is a 2 Àb 6 Àb 1 6 hUM (n, m) ¼ 6 Àb 25 6 4 Àb Àb (2:122)

simple 5 Â 5 moving average ﬁlter, then 3 Àb Àb Àb Àb Àb Àb Àb Àb 7 7 Àb 24 þ 25b Àb Àb 7 (2:123) 7 Àb Àb Àb Àb 5 Àb Àb Àb Àb with b ¼ 25 ¼ 2:2727 and 11 3 À1 À1 À1 À1 7 7 À1 À1 7 À1 À1 5 À1 À1

The following two unsharp masked ﬁlters are designed b ¼ 25 ¼ 6:25. 4 2 À1 À1 À1 6 À1 À1 À1 1 6 hUM (n, m)jb¼2:2727 ¼ 6 À1 À1 35 11 4 À1 À1 À1 À1 À1 À1 and À1 6 À1 16 ¼ 6 À1 4 4 À1 À1 2

(2:124)

hUM (n, m)jb¼6:25

3 À1 À1 À1 À1 À1 À1 À1 À1 7 7 À1 28 À1 À1 7 À1 À1 À1 À1 5 À1 À1 À1 À1

(2:125)

In general, higher values of b result in more image sharpening. Similarly, in the above ﬁlters, a smaller center tap implies more sharpening since it increases the relative contribution of the negative taps on the ﬁltered output.

2.8.2 IMAGE HISTOGRAM

Consider an M Â M, 8 bit image f (n, m). The image histogram N(f ), for each code value f is the number of times the code value f appears in the image. The PDF of image f (n, m) is the normalized histogram deﬁned by p(f ) ¼ N(f ) M2 f ¼ 0, 1, . . . , 255 (2:126)

Fundamentals of Digital Image Processing

79

**The cumulative distribution function CDF(f ) is deﬁned as CDF(f ) ¼
**

f X i¼0

p(i)

(2:127)

As an example, the histogram, PDF and CDF for the LENA image are shown in Figures 2.43 through 2.45.

1000

800

600 N( f ) 400 200 0 0

50

100 150 Code value f

200

250

**FIGURE 2.43 Histogram of LENA image.
**

0.015

0.01 P( f ) 0.005 0 0

50

100

150

200

250

Code value f

FIGURE 2.44 PDF of LENA Image.

80

Control of Color Imaging Systems: Analysis and Design

1

0.8

CDF( f )

0.6

0.4

0.2

0

0

50

100 150 Code value f

200

250

FIGURE 2.45 CDF of LENA image.

2.8.3 HISTOGRAM EQUALIZATION

Assume that r is a continuous random variable taking values between 0 and 1 with a PDF pr (r). Let T(r) be a single-valued and monotonically increasing function of its argument such that 0 T(r) 1 for 0 r 1 (2:128)

Then, according to the theory of functions of one random variable, the random variable s ¼ T(r) has a PDF given by ps (s) ¼ pr (r)

dT dr

jr¼T À1 (s)

(2:129)

Now if we want the random variable s to be uniformly distributed over the interval [0 1], then we set ps (s) ¼ 1. Substituting this into Equation 2.129 yields 1¼ or dT ¼ pr (r) dr Integrating both sides of Equation 2.131, we have

r ð s ¼ T(r) ¼ pr (x)dx 0

pr (r)

dT dr

(2:130)

(2:131)

(2:132)

Fundamentals of Digital Image Processing

81

Example 2.18

Image f with gray levels between 0 and 1 has a PDF as shown in Figure 2.46. Find the transformation that changes the image f to a new image having a uniform PDF. The transformation that changes the PDF of f to a uniform distribution is given by Equation 2.133.

r ð r ð

s ¼ T(r) ¼ pr (x)dx ¼ 2xdx ¼ r2

0 0

(2:133)

This means that each gray level in the output image is the square of the corresponding gray level in the input image. In histogram equalization, the objective is to obtain a uniform histogram for the output image. The transformation that maps the input image to the output image is obtained by discrete approximation to the integral Equation 2.133. Assuming that the input image has L levels, then s¼ CDF(r) À a (L À 1) þ 0:5 1Àa ! r ¼ 0, 1, . . . , L À 1 (2:134)

where CDF(r) is the cumulative distribution function of the input image, a is the minimum value of CDF(r) and [x] stands for the largest integer less than or equal to x.

Example 2.19

The low contrast LENA image and its histogram are shown in Figure 2.47. The equalized image and its histogram are shown in Figure 2.48.

pr(r)

2

1

r

FIGURE 2.46 PDF of image f .

82

Control of Color Imaging Systems: Analysis and Design

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

FIGURE 2.47 Low contrast LENA image and its histogram.

0

0.2

0.4

0.6

0.8

1

FIGURE 2.48

Equalized image and its histogram.

2.9 IMAGE RESTORATION

The goal in image restoration is to reconstruct the original image which has been subjected to degradation and noise [21,22]. It is usually done by compensating for the image noise and blur, assuming certain degradation models. The closeness of the restored image to the original image depends on the accuracy of the model and the availability of the degradation operator and additive noise statistics. Wiener ﬁltering, weighted least square, and maximum likelihood are common techniques for image restoration. Figure 2.49 shows the block diagram of an image degradation and restoration system.

vy ) (2:137) In Equation 2. m) À ^(n. vy ) is the power spectrum of the original image. vy ) is the blur (degradation kernel) frequency response.137. m) * h(n. If there is no additive noise (SN (vx . SN (vx . m) ¼ f (n.Fundamentals of Digital Image Processing 83 f (n.49 Image degradation and restoration ﬁlter. 2. vy ) ¼ H(vx . W(vx . then the Wiener ﬁlter becomes H*(vx . m) is the degraded noisy image. vy ) W(vx . The solution to the above minimization problem can be obtained using calculus of variation and is given by H*(vx . The restoration ﬁlter is designed by minimf izing a metric that is a measure of closeness of ^ to f according to some criterion. m) Restoration filter fˆ(n. vy ) Sf (vx . m) ¼ f (n. vy ) (2:138) . H(vx . m) h(n. This is known as MMSE ﬁlter or Wiener Filter. vy ) H(vx . vy ) ¼ 2 ¼ H(vx . h(n. m)j2 ) (2:136) where e(n. m) (2:135) where f (n. we discuss an approach based on minimizing the mean-square error between the original image and the restored image. m) + + Σ N(n. m) is the restored image. vy ) 1 W(vx . g(n. m) Degradation model FIGURE 2. N(n.1 WIENER FILTER RESTORATION The Wiener ﬁlter is designed by minimizing the following objective function J ¼ E(je(n. vy ) is the frequency response of the Wiener ﬁlter (restoration ﬁlter). m) is the original image. vy ) ¼ 0). The observed image g(n. vy )2 þ SN (vx . m) is given by g(n. vy ) is the noise power spectrum and Sf (vx .9. m) is a linear ﬁlter representing the blur and ^(n. m) is zero mean stationary additive noise process. m) g(n. m) þ N(n. f Here. m) is the estimation error and E is the expected value f operator.

If there is no degradation except for additive noise. then W(vx . v ) spectra are not known and the ratio SN(vxx. m) g g(n. vy ) (2:139) This is the Wiener ﬁlter for noise removal. This is called inverse ﬁlter. The window w(n. vy ) ¼ 1.50 FFT implementation of the Wiener ﬁlter.50. vy ) Sf (vx . In practice. vy ) ¼ H(vx . H(vx . It is generally chosen by trial and error.84 Control of Color Imaging Systems: Analysis and Design N M M ˆ (n. m) W(ωx .51. m) is a separable 2-D window given as product of two 1-D windows. vy ) W(vx . The original image is shown in Figure 2. The Wiener ﬁlter is normally implemented in frequency domain using the FFT algorithm. m) ¼ w1 (n)w2 (m) (2:141) Example 2.50. m) Zero padding 0 0 ˆ g (n. the noise and the image power S (v .m) 0 N N×N FFT w(n. . vyy) is replaced by the constant a.m) N N×N IFFT FIGURE 2. w(n. As indicated in Figure 2. vy ) ¼ 1 1þ SN (vx . the blur frequency response is sampled and multiplied by the FFT of the zero-padded and windowed input image. vy )2 þ a (2:140) The constant a is a measure of noise power to image signal power (inverse of SNR).20 The LENA image is blurred by a 5 Â 5 moving average ﬁlter and white Gaussian noise is added to the resulting image. f H*(vx . ωy) Uniform sampling over N × N grid N fˆ(n. that is. The block diagram of FFT-based implementation of the Wiener ﬁlter is shown in Figure 2.

the inverse ﬁlter ampliﬁes the noise to an extent that the image is not visible.53. Halftone images appear as if they are gray scale images. This is due to local .52 Blurred noisy LENA image. Most of the printed images in this book are halftone images. The Wiener ﬁlter reduces the noise effect and the resulting image is closer to the original image than the inverse ﬁltered image. The restored image using the inverse ﬁlter is shown in Figure 2.54. FIGURE 2. 2.10 IMAGE HALFTONING Digital image halftoning converts a gray scale image to a binary image for display and printing on binary devices. As shown.52. The blurred and noisy image is shown in Figure 2.Fundamentals of Digital Image Processing 85 FIGURE 2. The best result is obtained when a ¼ 0:01 in the Wiener ﬁlter.51 Original LENA image. The restored image using the Wiener ﬁlter is shown in Figure 2.

f (n.55. A simple technique for generating a halftone image is shown in Figure 2. m) Σ 1-Bit quantizer g(n.86 Control of Color Imaging Systems: Analysis and Design FIGURE 2. FIGURE 2. m) is added to each continuous tone image sample f (n. m) + + y(n. A halftone screen value s(n. spatial averaging (low-pass ﬁltering) performed by the eye. m) Halftone image Halftone screen s(n.54 Restored image using the Wiener ﬁlter. m). The resulting signal is denoted by y(n. m) FIGURE 2.53 Restored image using inverse ﬁlter.55 Halftone image generation. m) and is given by .

52 6 60 6 6 68 6 6 76 6 Halftone pattern ¼ 6 6 132 6 6 200 6 4 212 204 2 44 4 12 84 140 228 252 196 36 28 20 92 148 236 244 188 124 116 108 100 156 164 172 180 132 200 212 204 52 60 68 76 140 228 252 196 44 4 12 84 148 236 244 188 36 28 20 92 3 156 164 7 7 172 7 7 180 7 7 7 124 7 7 116 7 7 108 5 100 (2:144) is an example. m) to 1 bit.1. that is. Equation 2.57 and is generated using the halftone pattern shown in Equation 2. m) is shown in Figure 2.55. m) The halftone image is obtained by quantizing y(n. m) ¼ 1 0 y(n. m). the 512 Â 512 LENA image shown in Figure 2.Fundamentals of Digital Image Processing 87 y(n.56 is the input to the halftone algorithm shown in Figure 2.144 is a typical halftone pattern.10. FIGURE 2. & g(n. The resulting halftone image is shown in Figure 2. m) þ s(n. m) ! T otherwise (2:142) (2:143) The threshold T for an 8 bit image is 126. The halftone screen s(n. . This halftone pattern is repeated periodically to generate a halftone screen s(n. Another common halftoning technique is error diffusion technique discussed in Section 2. m) ¼ f (n. m) of the same size as the original continuous tone image f (n.58.144.56 Original image.

m) À y(n.88 Control of Color Imaging Systems: Analysis and Design FIGURE 2. m) to 1 bit. FIGURE 2. 2. m) À j i (2:145) where e(n.59. m) ¼ g(n.1 ERROR DIFFUSION ALGORITHM The block diagram of error diffusion algorithm is shown in Figure 2.58 Halftone image. m) XX h(i. m) ¼ f (n. m) is diffused to the neighboring pixels by the LPF h(n. m) À e(n. m) * h(n. The signal y(n. . m À j) ¼ f (n. The quantization error signal e(n. j)e(n À i.57 Halftone screen.10. m). m) is given as y(n. The halftone image is obtained by quantizing y(n. m).

61.147. The original image and the halftone image are shown in Figures 2.21 The gray scale LENA image is halftoned using the error diffusion algorithm with the ﬁlter given in Equation 2. The halftone image g(n. respectively.147. m) Low-pass filter FIGURE 2.59 Error diffusion halftoning. m) ! 126 otherwise (2:146) An example of a low-pass diffusion ﬁlter commonly used is 1 0 h¼ 16 3 0 5 7 1 ! (2:147) As indicated by the diffusion ﬁlter. the quantization error is diffused to the four neighboring pixels.60 and 2.62 and 2. The original image and the halftone image are shown in Figures 2. m) + – e (n. m) is generated by 1 bit quantization of y(n. m) h(n. m) y (n. m) ¼ 1 0 y(n.63. Example 2. Example 2. m) + – 1-Bit quantizer g (n.22 The RGB color LENA image is halftoned using the error diffusion algorithm with the ﬁlter given in Equation 2. respectively.Fundamentals of Digital Image Processing 89 f (n. . m) and is given by & g(n.

FIGURE 2.90 Control of Color Imaging Systems: Analysis and Design FIGURE 2.62 (See color insert following page 428. . FIGURE 2.) RGB LENA image.61 Halftone LENA image.60 LENA Image.

Find the radius of blur in terms of the number of pixels of the CCD sensor for the above set of parameters. respectively. and a CCD sensor pixel dimension of 16 mm Â 16 mm.4 A circularly symmetric imaging system has a PSF given by h(r) ¼ eÀjrj a. N À 1 2.2 Consider an out-of-focus camera with a blur radius of R.3 Find the DFT of the checkerboard image deﬁned by f (n. Repeat part (b) for an f number of 2 and an f number of 16. 1. . At what spatial radial frequency does the system MTF fall to pﬃﬃ? 2 . a. c. m ¼ 0. Assume that the object moves horizontally with uniform motion during the exposure period such that the resulting PSF on the ﬁlm has a width of a. m) ¼ 0:5 þ 0:5(À1)nþm n. 2.6. Find the horizontal proﬁle of the resulting cascaded PSF for a given vertical distance of y on the ﬁlm. . PROBLEMS 2. an f number of 5. Derive an expression for the OTF and MTF of this imaging system.) Halftone LENA image. 2.1 Consider a digital camera with a focal length of 50 mm. Is this distance halfway in between the two objects? b. . We wish to take a picture of these two objects and keep both of them in focus as much as possible. . 1 b.Fundamentals of Digital Image Processing 91 FIGURE 2. Consider two objects at distances of 1 and 3 m from the camera.63 (See color insert following page 428. Find the distance at which the camera should be focused so that the PSF resulting from the defocus of both of these objects have the same radius.

5 Consider a linear. y À y0 (t))dt 0 where g(x. vy ) ¼ eÀj(vx x0 (t)þvy y0 (t)) dt 0 b.64. Sketch the resulting MTF. we have 1 g(x. shift-invariant image degradation system with a PSF h(x. The total exposure at any point of the recording medium (e.g.7 A PSF has no spatial frequencies greater than 400 cycles=mm. a.64 2. y) undergoes planer motion. What values would you assign to the sampling interval Dx and the DFT length so as to obtain samples of the MTF in which aliasing is negligible and samples are spaced no further than 5 cycles=mm apart? The DFT length must be a power of 2. respectively. at the rate of x0 (t) ¼ 0:5at 2 . 2. What is the output image g(x..6 Suppose that an image f (x.65. and let x0 (t) and y0 (t) be the time-varying components of motion in the x. .and y-directions. y) ¼ eÀ2[jxjþjyj] . y) ¼ T T ð f (x À x0 (t). Then. Suppose the input to the system is an image consisting of one line as shown in Figure 2.92 Control of Color Imaging Systems: Analysis and Design x y FIGURE 2. y)? 2. 2. y) is the output image. if T is the duration of exposure.8 Find the Fourier transform of a line oriented at an angle u as shown in Figure 2. ﬁlm) is obtained by integrating the instantaneous exposure over the time interval during which the shutter is open. Suppose that the image undergoes motion in the x-direction only. Show that the OTF of the system is 1 T T ð H(vx .

66 n2 (1) x(n1. Develop a procedure to compute y(n1 . n2 ) * h(n1 .67. n2 ) ¼ x(n1 .Fundamentals of Digital Image Processing y 93 f (x. n2 ). .67 2. Determine y(n1 . n2 ) using DFT. n2 ) * h(n1 . The image is shown in Figure 2. a.10 Consider the two sequences x(n1 . n2) n2 h(n1. 2. Display the transform as an image. b.65 y 1 x FIGURE 2. n2 ). n2 ) shown in Figure 2. the linear convolution of x(n1 .66. n2 ) and h(n1 . n2) (2) (1) (4) (3) (2) n1 (2) (2) (2) (2) n1 FIGURE 2. y) θ x0 x FIGURE 2. n2 ) and h(n1 . y) ¼ d(x2 þ y2 À 1). n2 ) ¼ x(n1 .9 Find the Fourier transform of the 2-D image f (x.

m) to be L Â L. m) are separable.15 Let P(x) ¼ xeÀ0:5x u(x) and let the number of quantizer levels be 8 (3 bit). b.14 The signal f (x) is band limited so that its spectrum. a. What are the decision and reconstruction levels? b. m) ¼ 6 43 2 2 3 7 0 0 2 1 5 2 3 3 07 7 and 85 6 ! h(n. Show that the coefﬁcients. is zero outside the interval À2pB < v < 2pB. F(v) (i. The number of real multiplications if the convolution is performed in the pixel domain. m) ¼ 4 2 12 6 2. in this expansion are proportional to samples 1 of f (x) at the interval of 2B unit of length. m) to be N Â N and h(n. Assume f (n. a.e. . then. m) and h(n. Both f (n. m) is separable. Fourier transform). m) are nonseparable. m) with h(n. actually a statement of the most popular form of the sampling theorem.12 Determine the convolution of fx (n. 2.13 An image of size 64 Â 64 is to be ﬁltered using a 5 Â 5 FIR LPF. where 1 64 x(n. 2. m) for the three cases given below. m) is not separable but h(n. m). Expand the spectrum F(v) in a Fourier series on the interval À2pB < v < 2pB.68. a. in that it expresses the band-limited signal f (x) in terms of 1 discrete samples taken at intervals of 2B units of length. 2 2. Both f (n. having the form F(v) ¼ kv fk exp Àj 2B k¼À1 1 X b. Using these values for fk in the above equation. c. m) and h(n. 2. f (n.11 Determine the approximate number of multiplications required for the 2-D convolution of f (n.94 Control of Color Imaging Systems: Analysis and Design 2. fk . take the inverse Fourier transform of F(v) to give f (x) and show that the result expresses the signal f (x) in terms of the sample values k f ( 2B ) and an ‘‘interpolation function’’ of the form À Á k sin 2pB x À 2B À Á k 2pB x À 2B This last result is. m) and h(n. Find the resulting MSE..16 Design a 2-D 11 Â 11 zero-phase FIR ﬁlter to approximate the desired frequency response shown in Figure 2. Find a. The number of real multiplications if the overlap–add technique is used with an 8-point FFT algorithm. b.

5π D D ω1 0. for a nonseparable ﬁlter with the following frequency response.17 Consider the 8 Â 8 image f (n. 123 6 128 6 6 6 24 6 6 129 f (n. m) shown below. Assume this image is ﬁltered by a separable ﬁlter having the following horizontal and vertical frequency responses Hh (vx ) ¼ 1 (1 þ 2 cos (vx )). m) ¼ 6 6 190 6 6 6 100 6 4 86 23 2 125 129 150 180 185 23 12 14 130 135 120 109 98 88 0 4 123 124 140 150 100 136 5 8 140 145 123 155 162 123 135 128 145 140 134 160 89 35 34 56 150 129 29 29 68 187 45 25 3 15 19 7 7 7 35 7 7 65 7 7 34 7 7 7 43 7 7 56 5 23 a. v2 ) 2 D otherwise 2. v2 ) ¼ 1 0 (v1 . vy ) ¼ 4 À cos (vx ) À cos (vy ) .68 & Hd (v1 .5π 95 –0. 5) b. 3 Hv (vy ) ¼ 3 À 2 cos (vy ) What are the gray level values of the output image at the following locations? g(2. 3) and g(5.Fundamentals of Digital Image Processing ω2 0. H(vx .5π –0.5π FIGURE 2. Repeat part (1).

c. TABLE 2. b. 3 À4 À2 12 6 5 À4 À2 What is the DC gain of this ﬁlter (H(0.19 The histogram of a 32 Â 32. Determine a transformation that will achieve this objective.96 Control of Color Imaging Systems: Analysis and Design 2.20 Consider the following 3 bit gray level image.18 Consider the 3 Â 3 FIR ﬁlter 2 À2 14 h(n. m) ¼ 6 8 À2 a. 2. Find the histogram of this image. m) ¼ 6 62 6 65 6 42 2 2 4 1 2 6 1 1 0 3 3 2 5 6 5 3 7 1 3 3 5 2 2 3 0 3 1 1 7 4 3 4 6 4 1 3 3 3 1 1 7 2 0 2 2 3 4 2 5 4 3 0 67 7 37 7 67 7 37 7 77 7 75 3 a. 2 60 6 63 6 63 f (n. f (n.4 Histogram of Gray Scale Level Gray level Histogram 0 98 1 145 2 107 3 180 4 152 5 95 6 116 7 131 . b. m). d. e.4. What is the transfer function of the Wiener ﬁlter in terms of the power spectrum of the image? b. m) ! g(n. t 2. m) 2. Determine a transformation that will achieve this objective. We wish to modify the gray scale of this image such that the histogram of the processed image is as close as possible to being constant. Give an algorithm for FFT implementation of the Wiener ﬁlter. m) ¼ 2 þ cos (pn þ pm) ! h(n. We wish to modify the gray scale of this image such that the histogram of the processed image is as close as possible to being constant.21 A motion-blurred image (uniform motion in the x-direction with x0 (t) ¼ a T ) is observed in the presence of additive white noise with a power spectrum of Sn (vx . vy ) ¼ N0 . p)? Is this a zero-phase ﬁlter? Is it separable? Find the output image g(n. 3 bit gray level image is given in Table 2. a. 0))? What is the high-frequency gain H(p.

NJ. Lu. New York.A. R. Upper Saddle River. NJ. IRE Trans.. Discrete cosine transforms.E. Enlargement or reduction of digital images with minimum loss of information. New York. College Park. 76–79. Digital Image Processing: Concepts. Vector quantization. A new vector quantization clustering algorithm. IEEE Trans.W. 2001. Digital Image Processing. 17. 3. John Wiley & Sons. 2007. Springer-Verlag. An application of Heuristic search methods to edge and contour detection.. H. 1617–1627.K. 187–217. Graphics Image Proc. Springer-Verlag. IEEE Trans. R. 1999. Prentice Hall. University of Maryland. R. Unser. 1984. Info. 1990. New York. AU-15(2). and Rao. Acous. Technical Report TR-329. and Hildreth. 3rd ed. P. 1960. Histogram modiﬁcation techniques. . Digital Image Processing: Concepts. 11. 60. Review of imaging-blur models in a photographic system using the principles of optics. Edge detection using Heuristic search methods. 19. IT-6. Hummel. 4–29. The Fourier Transform and its Applications. Computer Science Center. The Fast Fourier Transform and its Applications. and Antoniou. 1995. 1989. NJ. M. Prentice Hall. P. 18. 4.-S. D.M...R. Comm. Lond. R. ASSP-37(10). Image Processing: The Fundamentals. 15.S. 3rd ed. Aldroubi. 1972. J. 7–12. 73–83.. Acous. A. 2. Image Processing. 5. B207. IEEE Trans. Upper Saddle River. 405–421. IEEE Trans.J. E. 9. 1985. J.O. 2000. Quantizing for minimum distortion.P. Clarke. 20. M. McGraw-Hill. J. and Welch.. Upper Saddle River. Natarajan.. Two-Dimensional Signal and Image Processing. Lim. Comput. Digital image restoration by constrained deconvolution. and Bosdogianni. Opt. Soc. 22.W. Lee. New York. Ahmed.. ACM. Two-Dimensional Digital Filters. Brigham. C-23.C. Prentice Hall. 10. Academic Press. Audio Electroacoustics. and Scientiﬁc Applications. Am.. 14. Equitz. 247–257. Jahne. R. Petrou. 1974.A. 16. Bracewell. 1970. 1974.Fundamentals of Digital Image Processing 97 REFERENCES 1.-C. Gray. A. 12. 90–93.. Comp. and Scientiﬁc Applications. A. J. 1. 6. Cooley. Historical notes on the fast Fourier transform. E. Speech Signal Processing.H. Martelli. K. 21. and Woods. 1988.. 1992. 1976. F-44620–72C-0062.K. 4(5). MD. Speech Signal Processing.N. Gonzalez. 1988. W. Pratt. A. P. 1568–1575. United Kingdom. Eng. Theory. 19(2). T. R. IEEE Trans. B. 29.D. Lewis. Martelli. J. W. Marcel Dekker. Transform Coding of Images. D. A. ASSP-1 (2). 1990. 1997.. Proc. W. R. New York. 169–182. Digital Image Processing. New York. and Eden. John Wiley & Sons. 1980. Algorithms. Opt. Jain. M. MacAdam. 13. Marr. 8. N. Theory of edge detection. Max. 1967.. Algorithms. 7. Soc.

.

Eigenvalues and eigenvectors. . .3 Mathematical Foundations 3. We will ﬁrst solve the equation in time domain.and discrete-time control systems. we introduce LTI continuous-time systems. Laplace transform. discrete-time systems.1 either in time domain or in the transform domain using Fourier or Laplace transform. and then we will use Laplace transform to solve the equation in s-domain. y(NÀ1) (0). the output signal y(t) can be determined uniquely by solving the DE given by Equation 3. The rest of this chapter is devoted to this important subject. Matrices and linear algebra are also important tools in analyzing control systems in the state-space form. y0 (0). 99 . A broad class of linear time-invariant (LTI) systems can be represented by linear differential equations (DEs) with constant coefﬁcients in case of continuous-time and linear difference equations with constant coefﬁcients in case of discrete-time systems. and functions of matrices are other mathematical tools used to design control systems with state-space approach.and closed-loop continuous.1 INTRODUCTION The motivation for this chapter is to present mathematical tools for analysis and design of open. 3. The Fourier and Laplace transforms play important roles in analyzing and designing such systems. and z-transform. singular value decomposition (SVD).2 GENERAL CONTINUOUS-TIME SYSTEM DESCRIPTION Consider a single-input single-output (SISO) continuous-time system with input u(t) and output y(t) that can be represented by constant-coefﬁcients linear DE of the form dN y(t) dNÀ1 y(t) dM u(t) dMÀ1 u(t) þ aNÀ1 þ Á Á Á þ a0 y(t) ¼ bM þ bMÀ1 þ Á Á Á þ b0 u(t) dt N dt NÀ1 dt M dt MÀ1 (3:1) where N is the order of the DE M is typically less than or equal to N Given the input signal u(t) and N initial conditions y(0). . . The z-transform is a tool for analyzing and designing discrete-time systems and is the counterpart of Laplace transform that is used for continuous-time systems. In this chapter.

These roots are generally complex and are called characteristic roots of the DE. the roots are D1 ¼ À1 and D2 ¼ À2. . y0 (0). that is. This implies that the characteristic roots must have real parts less than zero. . . . It is a polynomial of degree N that has N roots D1 . . then DN þ aNÀ1 DNÀ1 þ aNÀ2 DNÀ2 þ Á Á Á þ a0 ¼ 0 (3:4) (3:2) (3:3) This polynomial is called characteristic equation of the DE. . 2.1 SOLUTION CONSTANT-COEFFICIENTS LINEAR DIFFERENTIAL EQUATIONS We ﬁrst consider zero-input response that is the response of the system deﬁned by Equation 3. .2. . . . which can be written as D2 þ 3D þ 2 ¼ (D þ 1)(D þ 2) ¼ 0 Therefore. we would have Â Ã C DN þ aNÀ1 DNÀ1 þ aNÀ2 DNÀ2 þ Á Á Á þ a0 eDt ¼ 0 Since CeDt is nonzero. and DN . .1 is stable if the zero-input response decays to zero as t ! 1. . The stability of the system is deﬁned in terms of the zero-input response. . Consider an Nth-order constantcoefﬁcients linear DE with zero input driven by initial conditions y(0). C2 . The zero-input response is y(t) ¼ C1 eD1 t þ C2 eD2 t ¼ C1 eÀt þ C2 eÀ2t . Re(Di ) < 0 Example 3. . and CN are found by applying the N initial conditions. N (3:6) SOLUTION The characteristic equation is D2 þ 3D þ 2 ¼ 0.100 Control of Color Imaging Systems: Analysis and Design OF 3.1 when the input signal u(t) is zero. The zero-input solution is given by y(t) ¼ C1 eD1 t þ C2 eD2 t þ Á Á Á þ CN eDN t (3:5) The constants C1 . .1 Solve the following second-order DE with initial conditions y(0) ¼ À1 and dy(t) dt jt¼0 ¼ 4: d2 y(t) dy(t) þ3 þ 2y(t) ¼ 0 dt2 dt for i ¼ 1. D2 . and y(NÀ1) (0): dN y(t) dNÀ1 y(t) þ aNÀ1 þ Á Á Á þ a0 y(t) ¼ 0 N dt dt NÀ1 Assuming a solution of the form y(t) ¼ CeDt . . The discrete LTI system described by DE given in Equation 3.

7 yields 16AeÀ4t À 12AeÀ4t þ 2AeÀ4t ¼ eÀ4t Solving for A yields A¼1 6 Therefore.2 Solve the following second-order DE with initial conditions y(0) ¼ À1 and dy(t) dt jt¼0 ¼ 4: d2 y(t) dy(t) þ3 þ 2y(t) ¼ eÀ4t u(t) dt2 dt where u(t) is the unit step function. for example. yh (t) ¼ AeÀ4t Substituting Equation 3.8 into Equation 3. (3:7) SOLUTION Since the input is an exponential function. we consider the homogenous solution. y(t) ¼ 2eÀt À 3eÀ2t for t ! 0 101 Next. if the input signal is an exponential function. The homogenous solution is the result of driving the system with an input with zero initial conditions. Example 3. the response will be exponential and if the input signal is a sinusoid. the homogenous solution will be another exponential. This is due to the properties of linear systems. Therefore. the total solution is the sum of the zero-input response and the homogeneous solution y(t) ¼ C1 eÀt þ C2 eÀ2t þ 1 eÀ4t 6 (3:8) . A simple example that shows the process of obtaining the homogenous solution is given below. the output is also sinusoid with different amplitude and phase.Mathematical Foundations The constants C1 and C2 are found by applying the initial conditions y(0) ¼ C1 þ C2 ¼ À1 y0 (0) ¼ ÀC1 À 2C2 ¼ 4 The solutions are C1 ¼ 2 and C2 ¼ À3. The homogenous solution is typically similar to the input signal. that is.

9 is referred to as double-sided Laplace transform. 3 2 y(t) ¼ 7 eÀt À 7 eÀ2t þ 1 eÀ4t 3 2 6 for t ! 0 Another approach for solving constant-coefﬁcients linear DEs is use of Fourier or Laplace transform. This is due to the fact that in control systems. In the next section.3 LAPLACE TRANSFORM The Laplace transform of a continuous signal x(t) is a mapping from time domain to complex frequency domain deﬁned as 1 ð X(s) ¼ À1 x(t)eÀst dt (3:9) where s ¼ s þ jv is the complex frequency. SOLUTION 1 ð Àst 1 ð Àat Àst 1 ð X(s) ¼ 0 x(t)e dt ¼ 0 e e dt ¼ 0 eÀ(sþa)t dt ¼ À1 À(sþa)t 1 1 j0 ¼ e sþa sþa . we are mainly concerned with transient response of the system. is known as the region of convergence (ROC) of X(s). Therefore. 3.9 may not converge for all values of s. we brieﬂy cover Laplace transform with its properties. Interested readers can refer to Ref.3 Find the Laplace transform and ROC of the signal x(t) ¼ eÀat u(t). where the complex function X(s) converges. The transform deﬁned by Equation 3. Therefore. The integral in Equation 3. we are interested in one-sided Laplace transform deﬁned as 1 ð X(s) ¼ 0 x(t)eÀst dt (3:10) Example 3.102 Control of Color Imaging Systems: Analysis and Design The constants C1 and C2 are found by applying the initial conditions: y(0) ¼ C1 þ C2 þ 1 ¼ À1 6 y0 (0) ¼ ÀC1 À 2C2 À 4 ¼ 4 6 The solutions are C1 ¼ 7 and C2 ¼ À 7. [1] for more details. The region in complex plain s. the signals are deﬁned over the time interval of 0 to 1. In control applications.

1. . SOLUTION This is a special case of Example 3. where u(t) is the unit step function. as shown in Figure 3.5 Find the Laplace transform and ROC of the unit impulse signal x(t) ¼ d(t). the ROC is the entire complex s-plane. which is outside ROC of X(s). [1] for proofs of these properties. Example 3. The Laplace transform of some elementary signals and their ROCs are listed in Table 3.1 ROC of X(s).2. Thus.4 Find the Laplace transform and ROC of the signal x(t) ¼ u(t). Notice that the function X(s) has a single pole located at s ¼ Àa. the ROC is the set of points in the complex s-plane for which Re(s) > Àa. Properties of Laplace transform are given in Table 3.1. Interested readers are referred to Ref. SOLUTION 1 ð X(s) ¼ 0 x(t)eÀst dt ¼ 1 ð d(t)eÀst dt ¼ d(t)eÀst jt¼0 ¼ 1 0 Since the above integral converges for all values of s. Therefore. The convergence of X(s) requires that limt!1 eÀ(sþa)t ! 0.3 where a ¼ 0. X(s) ¼ 1 s ROC: Re(s) > 0 Example 3.Mathematical Foundations jω ROC 103 –a σ FIGURE 3.

inverse Laplace transform is used to ﬁnd the corresponding time-domain function. Once the Laplace transform of the output signal is determined. Interested readers are referred to Ref.1 INVERSE LAPLACE TRANSFORM An important application of Laplace transform is in analysis of LTI continuous systems. . X(s) ¼ PM Pk¼0 N k¼0 bk sk ak sk ¼ QQ PM k¼0 i¼1 bk s k (s À pi )mi (3:11) where P pi . .3. [2] for a complete coverage of this subject. . Q are the poles of X(s) or roots of polynomial N ak sk k¼1 mi is the multiplicity of the ith pole . Partial Fraction Expansion If X(s) is a rational function. we only consider partial-fraction approach. 2. .1 Laplace Transform of Elementary Functions Signal d(t) u(t) eÀat u(t) teÀat u(t) t n eÀat u(t) cos v0 tu(t) sin v0 tu(t) eÀat cos v0 tu(t) eÀat sin v0 tu(t) d(t À T) 1 1 s 1 sþa 1 (s þ a)2 n! (s þ a)nþ1 1 s2 þ v2 0 s s2 þ v2 0 1 (s þ a)2 þ v2 0 sþa (s þ a)2 þ v2 0 eÀsT Transform ROC s-Plane Re(s) > 0 Re(s) > ÀRe(a) Re(s) > ÀRe(a) Re(s) > ÀRe(a) Re(s) > 0 Re(s) > 0 Re(s) > ÀRe(a) Re(s) > ÀRe(a) s-Plane 3.104 Control of Color Imaging Systems: Analysis and Design TABLE 3. There are different techniques such as inversion integral and partialfraction expansion to compute the inverse Laplace transform from a given algebraic expression. In this section. that is. i ¼ 1. This analysis involves computing the response of the system to a given input using Laplace transform.

. . . Once the partial fraction is completed. . . . Assuming that N > M. miÀ1 . 1 (3:13) Similarly. 2. . .Mathematical Foundations 105 TABLE 3. partial fraction of X(s) yields ! A1 A2 Am1 X(s) ¼ þ þ ÁÁÁ þ s À p1 (s À p1 )2 (s À p1 )m1 ! B1 B2 Bm2 þ ÁÁÁ (3:12) þ þ ÁÁÁ þ þ s À p2 (s À p2 )2 (s À p2 )m2 The residues A1 . the other set of residues corresponding to other poles are computed.2 Properties of the Laplace Transform Property x(t) À X(s) ! y(t) À Y(s) ! Linearity Differentiation (ﬁrst derivative) Differentiation (second derivative) Multiplication by t Shift Multiplication by t n Scaling Integration 0 L L L ax(t) þ by(t) À aX(s) þ bY(s) ! dx(t) L À sX(s) À x(0) ! dt d2 x(t) L 2 À s X(s) À sx(0) À x0 (0) ! dt 2 tx(t) À À ! L L dX(s) ds ! x(t À t0 ) À eÀst0 X(s) tn eÀat u(t) À ! x(at) À ! t ð L L 1 X jaj a L n! (s þ a)nþ1 s x(t)dt À ! L X(s) s Convolution x(t) * y(t) À X(s)Y(s) ! Note that m1 þ m2 þ Á Á Á þ mQ ¼ N. the time-domain function x(t) is x(t) ¼ A1 ep1 t þ A2 ! tep1 t t 2 ep 1 t t m1 À1 ep1 t þ A3 þ Á Á Á þ Am1 1! 2! (m1 À 1)! þ B1 ep2 t þ B2 ! tep2 t t 2 ep2 t t m2 À1 ep2 t þ B3 þ Á Á Á þ Bm2 þ ÁÁÁ 1! 2! (m2 À 1)! (3:14) . Am1 corresponding to the pole p1 are computed using Ak ¼ dkÀmi (s À p1 )m1 X(s)js¼p1 dzkÀmi k ¼ mi . A2 .

SOLUTION Taking Laplace transform from both sides of the above DE. we would have X(s) ¼ where s ¼ À2 (s þ 1)2 s ¼ À1 A2 ¼ lim (s þ 1)2 X(s) ¼ lim s!À1 s!À1 s þ 2 2 d(s þ 1) X(s) d s 2 ¼2 ¼ lim ¼ lim A1 ¼ lim s!À1 s!À1 ds s þ 2 s!À1 (s þ 2)2 ds B1 ¼ lim (s þ 2)X(s) ¼ lim s!À2 s!À2 s A1 A2 B1 þ ¼ þ (s þ 1)2 (s þ 2) s þ 1 (s þ 1)2 s þ 2 Therefore.7 Solve the following second-order DE using Laplace transform. dt d2 y(t) dy(t) þ3 þ 2y(t) ¼ eÀ4t u(t) dt2 dt where u(t) is the unit step function. The initial conditions are y(0) ¼ À1 and dy(t) jt¼0 ¼ 4.106 Control of Color Imaging Systems: Analysis and Design Example 3. X(s) ¼ Hence.6 Find the inverse Laplace transform of the following function of the right-sided signal x(t) if X(s) ¼ s (s þ 1)2 (s þ 2) SOLUTION Using partial-fraction expansion. we have s2 Y(s) À sy(0) À y0 (0) þ 3[sY(s) À y(0)] þ 2Y(s) ¼ 1 sþ4 . x(t) ¼ [2eÀt À teÀt À 2eÀ2t ]u(t) 2 1 2 À À s þ 1 (s þ 1)2 s þ 2 Example 3.

This equation describes a general LTI discrete-time system. y(t) ¼ À7 Àt 3e Á À 7 eÀ2t þ 1 eÀ4t u(t) 2 6 3. The theory of z-transform with its properties and inverse z–transform are covered in the subsequent sections.4 GENERAL LINEAR DISCRETE-TIME SYSTEMS Consider a SISO discrete-time system with input u(n) and output y(n) that can be represented by DE of the form y(n) ¼ À N X i¼1 ai y(n À i) þ M X i¼0 bi u(n À i) (3:15) This deﬁnes an Nth-order constant-coefﬁcients DE. A2 . Y(s) ¼ Às2 À 3s þ 5 (s þ 4)(s2 þ 3s þ 2) 1 sþ4 107 Using partial-fraction expansion. . we have (s2 þ 3s þ 2)Y(s) þ s À 1 ¼ Therefore.Mathematical Foundations Using the initial conditions. In the next section. we solve the DE in time domain. This system can be analyzed in time domain or in transform domain using z-transform. and A3 are computed as A1 ¼ lim (s þ 1)Y(s) ¼ s!À1 Às2 À 3s þ 5 7 ¼ (s þ 2)(s þ 4) 3 Às2 À 3s þ 5 7 ¼À (s þ 1)(s þ 4) 2 Às2 À 3s þ 5 1 ¼ (s þ 1)(s þ 2) 6 A2 ¼ lim (s þ 2)Y(s) ¼ s!À2 A3 ¼ lim (s þ 4)Y(s) ¼ s!À4 Therefore. we have Y(s) ¼ Y(s) ¼ Às2 À 3s þ 5 Às2 À 3s þ 5 ¼ 2 þ 3s þ 2) (s þ 4)(s (s þ 4)(s þ 1)(s þ 2) A1 A2 A3 þ þ sþ1 sþ2 sþ4 The residues A1 .

jDi j < 1 for i ¼ 1. . .4. The stability of the system is deﬁned in terms of the zero-input response. we would have CDn ¼ À or CD Since CDn is nonzero. . that is. .15 when the input signal u(n) is zero. . and DN .8 Solve the following second-order DE with initial conditions y(À2) ¼ 1 and y(À1) ¼ À1: y(n) ¼ 0:75y(n À 1) À 0:125y(n À 2) (3:22) . . . . . y(À2). The zero-input solution is given by y(n) ¼ C1 Dn þ C2 Dn þ Á Á Á þ CN Dn 1 2 N (3:20) The constants C1 . then 1þ N X i¼1 n N X i¼1 ai CDnÀi (3:17) 1þ N X i¼1 ! ai D Ài ¼0 (3:18) ai DÀi ¼ 0 (3:19) This is called characteristic equation of the DE. and CN are found by applying the N initial conditions. This implies that the characteristic roots must have magnitude less than one. .15 is stable if the zeroinput response decays to zero as n ! 1. D2 . These roots are generally complex and are called characteristic roots of the DE. 2. C2 . and y(ÀN): y(n) ¼ À N X i¼1 ai y(n À i) (3:16) Assuming a solution of the form y(n) ¼ CDn .108 Control of Color Imaging Systems: Analysis and Design OF 3.1 SOLUTION CONSTANT-COEFFICIENTS DIFFERENCE EQUATIONS We ﬁrst consider zero-input response that is the response of a system deﬁned by Equation 3. It is a polynomial of degree N that has N roots D1 . . . N (3:21) Example 3. . Consider an Nth-order constant-coefﬁcients DE with zero input driven by initial conditions y(À1). . . The discrete LTI system described by DE given in Equation 3. .

9 Consider the following second-order DE: y(n) ¼ 0:75y(n À 1) À 0:125y(n À 2) þ u(n) (3:28) Find the output signal y(n) if the input signal is u(n) ¼ 2(0:4)n for n ! 0. that is. Assume zero initial conditions. that is. Therefore.29 into Equation 3. the response will be exponential. the roots are D1 ¼ 0:5 and D2 ¼ 0:25. if the input is an exponential function. Example 3. we have A(0:4)n ¼ [0:75(0:4)À1 A À 0:125(0:4)À2 A þ 2](0:4)n (3:31) (3:30) (3:29) . The homogenous solution is found by driving the system with an input but zero initial conditions. the homogenous solution will be another exponential.28 yields A(0:4)n ¼ 0:75A(0:4)nÀ1 À 0:125A(0:4)nÀ2 þ 2(0:4)n Simplifying the right side of Equation 3.30. SOLUTION Since the input is an exponential function. The homogenous solution is similar to input signal. yh (n) ¼ A(0:4)n Substituting Equation 3.Mathematical Foundations 109 SOLUTION The characteristic equation 1 À 0:75DÀ1 þ 0:125DÀ2 ¼ 0 can be written as D2 À 0:75D þ 0:125 ¼ (D À 0:5)(D À 0:25) ¼ 0 (3:23) Therefore. y(n) ¼ À1:25(0:5)n þ 0:375(0:25)n n ! À2 (3:27) (3:25) (3:26) (3:24) Next. The zero-input response is y(n) ¼ C1 (0:5)n þ C2 (0:25)n The constants C1 and C2 are found by applying the initial conditions y(À1) ¼ C1 (0:5)À1 þ C2 (0:25)À1 ¼ 2C1 þ 4C2 ¼ À1 y(À2) ¼ C1 (0:5)À2 þ C2 (0:25)À2 ¼ 4C1 þ 16C2 ¼ 1 The solutions are C1 ¼ À1:25 and C2 ¼ 0:375. we consider the homogenous solution. for example. if the input is a sinusoid. let y(À2) ¼ y(À1) ¼ 0. the output is also a sinusoid.

we deﬁne z-transform. and properties associated with the transform. y(n) ¼ C1 (0:5)n þ C2 (0:25)n þ yh (n) ¼ C1 (0:5)n þ C2 (0:25)n À 64 (0:4)n 3 (3:35) The constants C1 and C2 are found by applying the initial conditions y(À1) ¼ C1 (0:5)À1 þ C2 (0:25)À1 À y(À2) ¼ C1 (0:5)À2 þ C2 (0:25)À2 À 160 160 ¼ 2C1 þ 4C2 À ¼0 3 3 400 400 ¼ 4C1 þ 16C2 À ¼0 3 3 (3:36) (3:37) Solving Equations 3. 3 y(n) ¼ 20(0:5)n þ 10 64 (0:25)n À (0:4)n 3 3 n ! À2 (3:38) DE can also be solved using z-transform.110 Therefore.39 may not converge for all values of z. We will then use z-transform to solve discrete-time linear systems. In the next section. The inﬁnite sum in Equation 3. yh (n) ¼ À 64 (0:4)n 3 n!0 2 64 ¼À À2 3 1 À 0:75(0:4) þ 0:125(0:4) À1 (3:32) (3:33) (3:34) The total solution is the sum of the homogenous and the zero-input response. Therefore. If the signal x(n) ¼ 0 for n < 0.36 and 3. Therefore. we have C1 ¼ 20 and C2 ¼ 10. then we have one-sided z-transform.37. Control of Color Imaging Systems: Analysis and Design A ¼ 0:75(0:4)À1 A À 0:125(0:4)À2 A þ 2 Solving for A yields A¼ Therefore. inverse z-transform. The region where the complex function X(z) converges is known as the region of convergence. 3. This is referred to as double-sided z-transform. which is deﬁned as .5 z-TRANSFORM The z-transform of discrete-time signal x(n) is deﬁned as X(z) ¼ 1 X n¼À1 x(n)zÀn (3:39) where z is a complex variable.

10 Find the z-transform of the one-sided signal x(n) ¼ an u(n).Mathematical Foundations 111 X(z) ¼ 1 X n¼0 x(n)zÀn (3:40) Convergence of the z-transform requires that 1 X n¼À1 jx(n)r Àn j < 1 (3:41) for some positive values of r. the ROC is the set of points in complex z-plane for which jzj > jaj. as shown in Figure 3.2 ROC for Example 3. Thus. This means absolute summability of the exponentially weighted sequence x(n).41. Notice that the function X(z) has a single pole located at z ¼ a.10. x1 (n) ¼ A x3 (n) ¼ a Example 3. For example. They are not satisfying the condition given above by Equation 3. the following sequences do not have z-transform. .2. Im(z) |a| Re(z) FIGURE 3. since neither of these sequences multiplied by r Àn would be absolutely summable for any value of r. sin (v0 n) v0 n x2 (n) ¼ A cos (v0 n þ u) n (3:42) (3:43) (3:44) À1 < n < 1 SOLUTION X(z) ¼ 1 X n¼À1 x(n)zÀn ¼ 1 X n¼0 an zÀn ¼ 1 X n¼0 (azÀ1 )n ¼ 1 z ¼ 1 À azÀ1 z À a The convergence of X(z) requires that jazÀ1 j < 1. which is outside ROC of X(z).

and zeros.11. any discrete-time sequence. where 0 < a < 1.12 Consider the sequence x(n) ¼ ajnj . as shown in Figure 3. which is the ratio of two polynomials in z (rational function).11 Consider the sequence x(n) ¼ cos (v0 n)u(n). . the sums in Equation 3. Example 3. These polynomials can be represented by a constant gain. which is the linear combination of exponential functions. Example 3. has a z-transform. In general.3 ROC for Example 3. The function X(z) has two complex conjugate poles at p1 ¼ ejv0 and p2 ¼ eÀjv0 . Find the z-transform of the sequence. The z-transform of this sequence is X(z) ¼ 1 X x(n)zÀn ¼ 1 X n¼0 cos (v0 n)zÀn ¼ 0:5 1 X n¼0 1 X n¼0 (ejnv0 þ eÀjnv0 )zÀn (3:45) ¼ 0:5 n¼À1 1 X n¼0 (ejv0 zÀ1 )n þ 0:5 (eÀjv0 zÀ1 )n If jejv0 zÀ1 j < 1 and jeÀjv0 zÀ1 j < 1 or jzj > 1.112 Control of Color Imaging Systems: Analysis and Design Im(z) e jω0 Poles of X(z) Re(z) e–jω0 FIGURE 3.45 converge and X(z) ¼ 0:5 1 1 0:5z 0:5z þ 0:5 ¼ þ 1 À ejv0 zÀ1 1 À eÀjv0 zÀ1 z À ejv0 z À eÀjv0 z(z À cos v0 ) ¼ 2 z À 2 cos v0 z þ 1 (3:46) The ROC of X(z) is the set of points outside the unit circle.3. a set of poles. These poles are outside ROC of X(z).

Mathematical Foundations Im(z) ROC of X(z) a O x a–1 x a < |z| < a–1 Re(z) 113 FIGURE 3. jzj < aÀ1 and jzj > a.47 must converge.4. both sums in Equation 3. that is. This requires that jazj < 1 and jazÀ1 j < 1. the ROC of X(z) is a < jzj < aÀ1 . if x1 (n) ! X1 (z). or equivalently.12. Z ROC ¼ R1 (3:48) . The pole–zero plot as well as ROC of X(z) is shown in Figure 3. Therefore. 3.4 Pole–zero plot and ROC of Example 3.5.1 PROPERTIES OF Z-TRANSFORM Most common properties of z-transform are listed below: (a) Linearity: z-Transform is a linear transform. SOLUTION X(z) ¼ 1 X n¼À1 x(n)zÀn ¼ À1 X n¼À1 (a)Àn zÀn þ 1 X n¼0 an zÀn ¼ À1 X n¼À1 (az)Àn þ 1 X n¼0 (azÀ1 )n This can be simpliﬁed as X(z) ¼ 1 X n¼1 (az)n þ 1 X n¼0 (azÀ1 )n ¼ À1 þ 1 X n¼0 (az)n þ 1 X n¼0 (azÀ1 )n (3:47) ¼ À1 þ or 1 1 þ 1 À az 1 À azÀ1 X(z) ¼ z(z þ a À aÀ1 ) (z À a)(z À aÀ1 ) For convergence of X(z). The z-transforms of some elementary functions and their ROCs are listed in Table 3.3.

114 Control of Color Imaging Systems: Analysis and Design TABLE 3. then ax1 (n) þ bx2 (n) ! aX1 (z) þ bX2 (z). Z Z ROC ¼ R2 (3:49) ROC ¼ R ¼ R1 \ R2 (3:50) This means that the z-transform of linear combination of two sequences is the linear combination of their respective z-transforms and the ROC is .3 z-Transform of Elementary Sequences x(n) d(n) d(n À k) u(n) an u(n) an u(Àn À 1) ajnj nu(n) nan u(n) n2 an u(n) n(n À 1) Á Á Á (n À m þ 1) nÀm a u(n À m þ 1) m! cos (v0 n)u(n) sin (v0 n)u(n) an cos (v0 n)u(n) an sin (v0 n)u(n) ajnj 1 z Àk X(z) ROC Entire z-plane Entire z-plane jzj > 1 jzj > jaj jzj < jaj a < jzj < jzj > 1 jzj > jaj jzj > jaj jzj > jaj jzj > 1 jzj > 1 jzj > jaj jzj > jaj a < jzj < 1 a 0<a<1 1 a 0<a<1 z zÀ1 z zÀa À z zÀa z(z þ a À aÀ1 ) (z À a)(z À aÀ1 ) z (z À 1)2 az (z À a)2 az(z þ a) (z À a)3 z (z À a)mþ1 z(z À cos v0 ) z2 À 2z cos v0 þ 1 z sin v0 z2 À 2z cos v0 þ 1 z(z À a cos v0 ) z2 À 2az cos v0 þ a2 za sin v0 z2 À 2za cos v0 þ a2 z(z þ a À aÀ1 ) (z À a)(z À aÀ1 ) and x2 (n) ! X2 (z).

The proof is straightforward and is based on the fact that sum of two sequences is equal to sum of individual sequences. we use the linearity property and the z-transform pair z an u(n) ! zÀa . Proof: 1 X n¼À1 ROC ¼ R (3:51) ROC ¼ R (3:52) x(n À k)u(n À k)zÀn ¼ 1 X n¼k x(n À k)zÀn ¼ 1 X l¼0 1 X l¼0 x(l)zÀlÀk (3:53) ¼ zÀk x(l)zÀl ¼ zÀk X(z) Example 3. Then X(z) ¼ 2 and the ROC is À Á À Á R ¼ ROC ¼ jzj > 1 \ jzj > 1 ¼ jzj > 1 3 2 2 z z z(5z À 2) ÁÀ Á þ3 ¼À zÀ1 zÀ1 zÀ1 zÀ1 3 2 3 2 (b) Delay Property of the z-Transform: If x(n)u(n) ! X(z). jzj > jaj. The delay k is an integer.13 Find the z-transform of the following sequence: x(n) ¼ 2 À1Án 3 À Án u(n) þ 3 1 u(n) 2 SOLUTION To ﬁnd X(z).14 Find the z-transform of the following sequence: & x(n) ¼ 0 (0:6)n n 3 n!4 . Example 3.Mathematical Foundations 115 the intersection of the individual regions of convergence. then x(n À k)u(n À k) ! zÀk X(z).

the z-transform of the output of a linear shift-invariant system is the product of the z-transform of the input signal and the z-transform of the impulse response of the system. The z-transform of the system impulse response h(n) is the system transfer function H(z). This means that the z-transform of convolution of two functions is the product of their corresponding z-transforms. If y(n) ¼ x(n) * h(n) then Y(z) ¼ X(z)H(z) Proof: Y(z) ¼ ¼ 1 X n¼À1 1 X k¼À1 (3:54) (3:55) y(n)zÀn ¼ x(k) 1 X n¼À1 1 X 1 X x(k)h(n À k)zÀn (3:56) n¼À1 k¼À1 h(n À k)zÀn Using shifty property. .116 Control of Color Imaging Systems: Analysis and Design SOLUTION The solution is easily found using the delay property of the z-transform. X(z) ¼ (0:6)4 zÀ4 z (0:6)4 ¼ 3 z À 0:6 z (z À 0:6) (c) Convolution Property: Convolution property of z-transform states that convolution in time domain is multiplication in z-domain. we have Y(z) ¼ 1 X k¼À1 x(k)H(z)zÀk ¼ H(z) 1 X k¼À1 x(k)zÀk ¼ H(z)X(z) (3:57) As a result of this important theorem. note that x(n) ¼ (0:6)n u(n À 4) ¼ (0:6)4 (0:6)nÀ4 u(n À 4) Therefore.

we have Y(z) ¼ X(z)H(z) ¼ (2 þ 3zÀ1 þ zÀ2 )(1 À 2zÀ1 þ 3zÀ2 þ 2zÀ3 ) ¼ 2 À zÀ1 þ zÀ2 þ 11zÀ3 þ 9zÀ4 þ 2zÀ5 Therefore.15 Convolve the following two sequences: x(n) ¼ ½ 2 3 1 and h(n) ¼ ½ 1 À2 3 2 SOLUTION The z-transforms of x(n) and h(n) are X(z) ¼ 2 þ 3zÀ1 þ zÀ2 H(z) ¼ 1 À 2zÀ1 þ 3zÀ2 þ 2zÀ3 Using the convolution property of z-transform.Mathematical Foundations 117 Example 3.16 Convolve the following two sequences using z-transform: x(n) ¼ (0:5)n u(n) h(n) ¼ (0:7)n u(n) SOLUTION The z-transform of y(n). the convolution of x(n) and h(n) is Y(z) ¼ X(z)H(z) ¼ z z z2 ¼ z À 0:5 z À 0:7 (z À 0:5)(z À 0:7) Performing partial fraction. we have Y(z) ¼ 3:5z 2:5z À z À 0:7 z À 0:5 Taking the inverse z-transform yields y(n) ¼ 3:5(0:7)n u(n) À 2:5(0:5)n u(n) . y(n) ¼ x(n) * h(n) ¼ ½ 2 À1 1 11 9 2 Example 3.

take the limit on both sides of Equation 3.58 with respect to the complex variable z results in 1 1 X X dX(z) ¼ x(n)(Àn)zÀnÀ1 ¼ ÀzÀ1 nx(n)zÀn dz n¼À1 n¼À1 (3:59) Multiplying both sides of Equation 3. where X(z) is the z-transform of x(n).118 Control of Color Imaging Systems: Analysis and Design (d) Multiplication by n or Differentiation of X(z): The z-transform of nx(n) is Àz dX(z). then the z-transform of x(n) ¼ nan u(n) is (3:62) d z zÀaÀz az ¼ Àz ¼ dz z À a (z À a)2 (z À a)2 (e) Initial Value Theorem: If x(n) ¼ 0 for n < 0. Example 3.17 Find the z-transform of the sequence x(n) ¼ nan u(n) (3:61) SOLUTION Since the z-transform of an u(n) is X(z) ¼ Àz z zÀa.59 by Àz yields Àz 1 X dX(z) ¼ nx(n)zÀn dz n¼À1 (3:60) The following example uses the differentiation property.64 as z approaches inﬁnity: X(z) ¼ 1 X n¼0 x(n)zÀn ¼ x(0) þ x(1)zÀ1 þ x(2)zÀ2 þ Á Á Á (3:64) . dz Proof: The z-transform of x(n) is X(z) ¼ 1 X n¼0 x(n)zÀn (3:58) Differentiating both sides of Equation 3. then x(0) ¼ lim X(z) z!1 (3:63) Proof: To derive this property.

then: lim (z À 1)X(z) ¼ x(1) z!1 (3:70) Example 3. we have x(0) ¼ lim X(z) ¼ lim z!1 z!1 z(z À 0:7) ¼1 z2 À 1:25z þ 0:25 . then x(1) ¼ lim (z À 1)X(z) z!1 (3:66) To derive this property. (z À 1)X(z) ¼ 1 X n¼À1 (3:67) [x(n þ 1) À x(n)]zÀn (3:68) Taking the limit as z ! 1 results in lim (z À 1)X(z) ¼ 1 X n¼À1 z!1 [x(n þ 1) À x(n)] (3:69) ¼ x(0) À x(À1) þ x(1) À x(0) þ x(2) À x(1) þ Á Á Á Since x(À1) ¼ 0.Mathematical Foundations 119 The result is lim X(z) ¼ x(0) þ 0 þ 0 þ Á Á Á ¼ x(0) (3:65) z!1 (f) Final Value Theorem: If x(n) ¼ 0 for n < 0 and X(z) does not have any poles on the boundary of unit circle. if X(z) ¼ z(z À 0:7) z2 À 1:25z þ 0:25 SOLUTION Using the initial and ﬁnal value theorems. we use the delay and linear properties of the z-transform.18 Find the initial and the ﬁnal values of the one-sided signal x(n). mainly (z À 1)X(z) ¼ zX(z) À X(z) ¼ z-transform of [x(n þ 1) À x(n)] Therefore.

we consider techniques such as inversion integral. . inverse z-transform is used to ﬁnd the corresponding time-domain sequence. In this section.2 INVERSE Z-TRANSFORM An important application of the z-transform is in the analysis of linear discrete-time systems. as shown in Figure 3. power series expansion. and partial fraction. we have x(n) ¼ X Residues of X(z)znÀ1 at poles of X(z) inside C (3:72) Im(z) Contour C a Re(z) b FIGURE 3. (a) Inversion Integral: The inverse z-transform of X(z) is given by the integral 1 x(n) ¼ j2p þ X(z)znÀ1 dz C (3:71) where C is any closed contour in the ROC of X(z) excluding the origin.120 and Control of Color Imaging Systems: Analysis and Design x(1) ¼ lim (z À 1)X(z) ¼ lim (z À 1) z!1 z!1 z(z À 0:7) z(z À 0:7) ¼ lim ¼ 0:4 (z À 1)(z À 0:25) z!1 z À 0:25 3. This analysis involves computing the response of the systems to a given input using z-transform. There are different techniques to ﬁnd the inverse z-transform from a given algebraic expression.5 Closed contour C. If X(z) is a rational function of its argument z. then using the residue theorem.5. Once the z-transform of the output signal is determined.5.

(z À 0:2)(z À 0:3) ROC: jzj > 0:3 SOLUTION The ROC and the contour C for this example are shown in Figure 3. .Mathematical Foundations 121 It can be shown that (P x(n) ¼ Residues of X(z)znÀ1 at poles of X(z) inside C P À Residues of X(z)znÀ1 at poles of X(z) outside C n!0 n À1 (3:73) Example 3.6. we have x(n) ¼ X Residues of z2 znÀ1 at poles of X(z) inside C ¼ A1 þ A2 (z À 0:2)(z À 0:3) The residue A1 corresponding to pole p1 ¼ 0:2 is A1 ¼ (z À 0:2) z2 0:22 (0:2)nÀ1 znÀ1 jz¼0:2 ¼ ¼ À2(0:2)n (z À 0:2)(z À 0:3) (0:2 À 0:3) Im(z) C 0.2 0.19. we have x(n) ¼ 0 for n < 0 For n ! 0.6 ROC and contour C for Example 3.19 Find the inverse z-transform of the following function: X(z) ¼ z2 . Since there are no poles outside C.3 Re(z) ROC of X(z) FIGURE 3.

If the function X(z) is a rational function. Â x(n) ¼ 1 0:2 (0:2)2 (0:2)3 (0:2)4 Ã Á Á Á ¼ (0:2)n u(n) Example 3. we expand X(z) as a P Àn power series of zÀ1 . the power series can be obtained using long division as illustrated by the following example. x(n) ¼ an . SOLUTION The power series expansion of X(z) is X(z) ¼ z 1 ¼ 1 þ 0:2zÀ1 þ (0:2)2 zÀ2 þ (0:2)3 zÀ3 þ Á Á Á ¼ z À 0:2 1 À 0:2zÀ1 Therefore. x(n) ¼ A1 þ A2 ¼ À2(0:2)n u(n) þ 3(0:3)n u(n) z2 0:32 (0:3)nÀ1 znÀ1 jz¼0:3 ¼ ¼ 3(0:3)n (z À 0:2)(z À 0:3) (0:3 À 0:2) (b) Inversion by Power Series Expansion: In this case. X(z) ¼ 1 n¼À1 an z .20 z Find the inverse z-transform of the function X(z) ¼ zÀ0:2 using power series expansion. that is. we have X(z) ¼ z3 À 2:3z2 þ 0:84z À 1:4z2 þ 0:63z À 0:09 ¼ 1 À 0:9zÀ1 þ 1:05zÀ2 À 0:813zÀ3 À 0:5577zÀ4 þ Á Á Á z3 .21 Find the inverse z-transform of X(z) ¼ z3 À 2:3z2 þ 0:84z z3 À 1:4z2 þ 0:63z À 0:09 SOLUTION Expanding X(z).122 Control of Color Imaging Systems: Analysis and Design and the residue A2 corresponding to the second pole is A2 ¼ (z À 0:3) Therefore. The expansion coefﬁcients are x(n). that is. Example 3.

Q are the poles of X(z) or roots of polynomial M ak zk k¼1 mi is the multiplicity of the ith pole. x(1) ¼ À0:9. we have log (1 þ azÀ1 ) ¼ 1 X n¼1 (À1)nþ1 an zÀn n Therefore. 2. .22 Find the inverse z-transform of the irrational function X(z) ¼ log (1 þ azÀ1 ) ROC: jzj > a SOLUTION Using Taylor series expansion. that is. . it is not possible to obtain closed-form solution using longdivision technique. . . x(n) ¼ (À1)nþ1 an u(n À 1) n (c) Inversion by Partial Fraction Expansion: If X(z) is a rational function. Example 3. . i ¼ 1. x(2) ¼ 1:05. .Mathematical Foundations Therefore. PM PM bk z k bk zk X(z) ¼ Pk¼0 ¼ QQ k¼0 M mi k k¼0 ak z i¼1 (z À pi ) (3:74) where P pi . 123 In general. x(0) ¼ 1. Note that m1 þ m2 þ Á Á Á þ mQ ¼ M Partial fraction of X(z) z yields ! X(z) A0 A1 A2 Am1 þ ¼ þ þ ÁÁÁ þ z z À p1 (z À p1 )2 (z À p1 )m1 z ! B1 B2 Bm2 þ þ ÁÁÁ þ þ ÁÁÁ þ z À p2 (z À p2 )2 (z À p2 )m2 (3:75) . . x(4) ¼ À0:5577. x(3) ¼ À0:813.

. Am1 corresponding to the pole p1 are computed using dkÀmi m1 X(z) Ak ¼ kÀm (z À p1 ) dz i z z¼p1 k ¼ mi . the other set of residues corresponding to other poles are computed. the time-domain function is x(n) ¼ A0 d(n) þ A1 (p1 )n u(n) þ A1 n(p1 )nÀ1 u(n) þ Á Á Á n(n À 1)(n À 2) Á Á Á (n À m1 þ 1) (p1 )nÀm1 u(n À m1 þ 1) þ Am1 m1 ! Â þ B1 (p2 )n u(n) þ B1 n(p2 )nÀ1 u(n) þ Á Á Á þ Bm2 ! ! n(n À 1)(n À 2) Á Á Á (n À m2 þ 1) (p2 )nÀm2 u(n À m2 þ 1) þ Á Á Á m2 ! (3:79) Example 3. Once the residues are computed. A2 . miÀ1 . .124 Control of Color Imaging Systems: Analysis and Design or X(z) ¼ A0 þ ! A1 z A2 z Am1 z þ þ ÁÁÁ þ z À p1 (z À p1 )2 (z À p1 )m1 ! B1 z B2 z Bm2 z þ þ ÁÁÁ þ þ ÁÁÁ þ z À p2 (z À p2 )2 (z À p2 )m2 (3:76) where X(z) ¼ X(0) A0 ¼ z z z¼0 (3:77) The residues A1 .23 Find the inverse z-transform of the following function: X(z) ¼ z2 . . . . . 2. 1 (3:78) Similarly. . we would have X(z) ¼ À30z 2z 30z þ À z À 0:2 (z À 0:2)2 z À 0:3 . (z À 0:2)2 (z À 0:3) ROC: jzj > 0:3 SOLUTION Using partial-fraction expansion. .

3 RELATION BETWEEN THE Z-TRANSFORM AND THE LAPLACE TRANSFORM 1 Consider analog signal xa (t) being sampled at the rate of fs ¼ T as shown in Figure 3. Y(z) ¼ z z ¼ (z2 þ 0:4z þ 0:03)(z À 1) (z þ þ0:1)(z þ 0:3)(z À 1) z zÀ1 Partial fraction of Y(z) results in z Y(z) 1 0:6993 4:5454 3:8461 ¼ ¼ À þ z (z þ þ0:1)(z þ 0:3)(z À 1) zÀ1 z þ 0:1 z þ 0:3 Y(z) ¼ Hence. y(n) ¼ [0:6993 À 4:5454(À0:1)n þ 3:8461(À0:3)n ]u(n) 0:6993z 4:5454z 3:8461z À þ zÀ1 z þ 0:1 z þ 0:3 3.24 Consider a SISO system described by the second-order DE: y(n þ 2) þ 0:4y(n þ 1) þ 0:03y(n) ¼ x(n) Find the output of this system if the input signal is a unit step function x(n) ¼ u(n) with the initial conditions y(1) ¼ y(0) ¼ 0. The output of the sampler could be considered to be either the discrete signal x(n) or the continuous signal x*(t) deﬁned as .7. SOLUTION Taking z-transform from both sides of the above DE yields z2 Y(z) þ 0:4zY(z) þ 0:03Y(z) ¼ X(z) ¼ Therefore.5.Mathematical Foundations Taking inverse z-transform yields x(n) ¼ À30(0:2)n u(n) À or x(n) ¼ [À30(0:2)n À 10n(0:2)n þ 30(0:3)n ]u(n) 2n(0:2)n u(n) þ 30(0:3)n u(n) 0:2 125 Example 3.

8. . the z-transform of a discrete signal x(n) is the Laplace transform of the sampled signal x*(t) with the change of variable z ¼ eTs (3:85) The above equation deﬁnes a mapping from complex s-plane to complex z-plane. we have X*(s) ¼ X(z)jz¼eTs (3:84) Thus.126 Control of Color Imaging Systems: Analysis and Design xa(t) x *(t) s(t) =n=–∞δ(t – nT) Σ ∞ FIGURE 3. as shown in Figure 3.82. Im(s) s-plane z = e sT Im(z) z-plane 1 Unit circle Re(z) Re(s) FIGURE 3. x(n) ¼ xa (nT) 1 X xa (nT)d(t À nT) x*(t) ¼ n¼À1 (3:80) (3:81) The Laplace transform of x*(t) is X*(s) ¼ 1 X n¼À1 xa (nT)eÀnTs (3:82) The z-transform of the discrete signal x(n) is X(z) ¼ 1 X n¼À1 x(n)zÀn (3:83) Comparing Equations 3.81 and 3.7 Sampling process.8 Mapping from the complex s-plane to the complex z-plane.

that is. that is.6 DISCRETE-TIME FOURIER TRANSFORM The discrete-time Fourier transform (DTFT) of discrete signal x(n) is deﬁned as X( jv) ¼ 1 X n¼À1 x(n)eÀjnv (3:86) This is the z-transform X(z) evaluated on the boundary of unit circle in z-plane. we brieﬂy cover some important properties of DTFT without proof. For this reason. X( jv) ¼ X[ j(v þ 2p)] (c) Delay Property: If x(n) X( jv) ! then x(n À n0 ) eÀjn0 v X( jv) ! Example 3.6. its properties are very similar to the properties of the z-transform. 3. if x(n) X( jv) ! y(n) Y( jv) ! then ax(n) þ by(n) aX( jv) þ bY( jv) ! (b) Periodicity: X( jv) is periodic with a period of 2p. discrete signal x(n) has a DTFT if the ROC of X(z) contains the unit circle. that is. Since the DTFT is closely related to the z-transform.25 Find the DTFT of the ﬁve-point discrete signal: x(n) ¼ ½ 2 À3 7 À3 2 DTFT DTFT DTFT DTFT DTFT (3:88) (3:89) (3:90) (3:91) (3:92) (3:93) . X( jv) ¼ X(z)jz¼ejv (3:87) Therefore.1 PROPERTIES OF DISCRETE-TIME FOURIER TRANSFORM (a) Linearity: DTFT is a linear transform.Mathematical Foundations 127 3.

that is. SOLUTION The DTFT of x(n) is X( jv) ¼ 1 X x(n)eÀjnv ¼ 2 À 3eÀjv þ 7eÀ2jv À 3eÀ3jv þ 2eÀ4jv (e2jv þ eÀ2jv ) À 3eÀ2jv (ejv þ eÀjv ) þ 7eÀ2jv ¼ 2e n¼À1 À2jv ¼ eÀ2jv [4 cos 2v À 6 cos v þ 7] The magnitude of X( jv) is jX( jv)j ¼ 7 À 6 cos v þ 4 cos 2v The magnitude plot is shown in Figure 3.9 Magnitude of DTFT of x(n). We can use this closed contour since x(n) has Fourier transform and its ROC contains the unit circle.2 INVERSE DTFT The inverse DTFT can be computed using the inversion integral used in computing the inverse z-transform. 1 x(n) ¼ j2p þ X(z)znÀ1 dz C (3:94) Let the integration contour C be the unit circle z ¼ ejv in the complex z-plane. 3.6.128 18 16 14 12 |X(ω)| 10 8 6 4 2 0 0 Control of Color Imaging Systems: Analysis and Design 1 2 3 ω 4 5 6 FIGURE 3. Therefore.9. .

26 Find the inverse DTFT of X( jv) given by X( jv) ¼ eÀjvj for Àp < v < p SOLUTION Using the inversion integral. z2 ) ¼ 1 X 1 X n1 ¼À1 n2 ¼À1 x(n1 .7 TWO-DIMENSIONAL z-TRANSFORM The two-dimensional (2-D) z-transform of the 2-D sequence x(n1 . we have 1 x(n) ¼ 2p p ð p ð X( jv)e Àp jnv 1 dv ¼ 2p eÀjvj ejnv dv Àp The above integral is decomposed into two integrals: 1 2p 0 ð x(n) ¼ ev ejnv dv þ Àp 1 2p p ð eÀv ejnv dv 0 Hence. n2 )zÀn1 zÀn2 1 2 (3:96) .Mathematical Foundations 129 p ð 1 x(n) ¼ j2p þ X(z)z C nÀ1 1 dz ¼ j2p X( jv)e Àp j(nÀ1)v 1 je dv ¼ 2p jv p ð X( jv)ejnv dv Àp (3:95) Example 3. 1 x(n) ¼ 2p ¼ 0 ð e Àp (1þjn)v 1 dv þ 2p p ð e 0 À(1Àjn)v À(1Àjn)v p 1 e(1þjn)v 0 À 1 e dv ¼ 2p 1 þ jn Àp 2p 1 À jn 0 1 1 À eÀ(1þjn)p 1 eÀ(1Àjn)p À 1 1 1 À eÀp eÀjnp 1 eÀp ejnp À 1 À À ¼ 1 þ jn 1 þ jn 2p 2p 1 À jn 2p 2p 1 À jn Therefore. n2 ) is deﬁned as X(z1 . x(n) ¼ 1 2 À 2eÀp cos np 1 À eÀp cos np ¼ 2p 1 þ n2 (1 þ n2 )p 3.

z2 ).23. jz1 j > jaj and jz2 j > jbj. The region in four-dimensional (4-D) {z1 .10. both sums in Equation 3. z2 ) ¼ 1 X 1 X n1 ¼À1 n2 ¼À1 x(n1 . z2 ) converges is deﬁned as ROC of X(z1 . X(z1 . Two-dimensional z-transform has applications in the design of 2-D inﬁnite impulseresponse ﬁlters and stability analysis of numerical solution of partial DEs. n2 ) ¼ an1 bn2 0 n1 . z2 ). n2 )zÀn1 zÀn2 ¼ 2 1 1 X n1 ¼0 (azÀ1 )n1 1 1 X n2 ¼0 (bzÀ1 )n2 2 (3:97) For convergence of X(z1 . z2 ) is shown in Figure 3. 1 2 Therefore. This requires that jazÀ1 j < 1 and jbzÀ1 j < 1 or equivalently.130 Control of Color Imaging Systems: Analysis and Design where z1 and z2 are complex variables. z2 ) ¼ 1 1 z1 z2 ¼ 1 À azÀ1 1 À bzÀ1 (z1 À a)(z2 À b) 2 1 The ROC of X(z1 . Example 3. It is typically shown as a set in the 2-D jz1 j – jz2 j plane. n2 ! 0 otherwise SOLUTION X(z1 . |z2| ROC |b| |z1| |a| FIGURE 3. The ROC depends on jz1 j and jz2 j.97 must converge.27 Find the z-transform of the sequence & x(n1 . z2 } space where X(z1 . .10 ROC of Example 3.

Mathematical Foundations |z2| 131 ROC |z1| FIGURE 3. n2 ) is X(z1 .28 Find the z-transform of the 2-D sequence x(n1 . z2 ).11 1 z1 z2 ¼ 1 À zÀ1 zÀ1 z1 z2 À 1 2 1 (3:100) 3.8 TWO-DIMENSIONAL DISCRETE-SPACE FOURIER TRANSFORM The 2-D discrete-space Fourier transform (DSFT) of 2-D discrete signal x(n1 .11 ROC of Example 3. Example 3. z2 ) evaluated on the boundary of unit circles z1 ¼ ejv1 and z2 ¼ ejv2 in the 4-D complex space (z1 . that is. jv2 ) ¼ X(z1 . jv2 ) ¼ 1 X 1 X x(n1 . X( jv1 . z2 ) ¼ 1 X 1 X 1 X n1 ¼0 n1 ¼À1 n2 ¼À1 x(n1 . z2 )jz1 ¼ejv1 .z2 ¼ejv2 (3:102) . z2 ) ¼ The ROC is shown in Figure 3. n2 ) is deﬁned as X( jv1 . n2 )zÀn1 zÀn2 ¼ 2 1 (z1 z2 )Àn1 (3:99) If jzÀ1 zÀ1 j < 1 or jz1 jjz2 j > 1.99 convergences to 1 2 X(z1 .24. n2 )eÀjn1 v1 eÀjn2 v2 (3:101) n2 ¼À1 n1 ¼À1 This is the z-transform X(z1 . n2 ) ¼ n 1 0 n1 ¼ n2 ! 0 otherwise (3:98) SOLUTION The 2-D z-transform of x(n1 . the sum in Equation 3.

12. jv2 ) ! (b) Periodicity: X( jv1 . n2 ) X( jv1 . jv2 ) is periodic with a period of 2p. Here are some important properties of 2-D DSFT: (a) Linearity: 2-D DSFT is a linear transform. n2 ) Y( jv1 . that is. jv2 ) ¼ X[ j(v1 þ 2p). n2 ) X( jv1 .8.2 INVERSE 2-D DSFT The inverse 2-D DSFT can be computed using the inversion integral given by 1 x(n1 .132 Control of Color Imaging Systems: Analysis and Design OF 3. jv2 ) ! DSFT DSFT DSFT DSFT DSFT (3:103) (3:104) (3:105) (3:106) (3:107) (3:108) 3.8. that is. jv2 ) ! y(n1 . if x(n1 . jv2 ) ¼ 1 0 if (v1 . X( jv1 . jv2 ) þ bY( jv1 . jv2 ) ! then x(n1 À a. n2 ) aX( jv1 . jv2 )ejn1 v1 ejn2 v2 dv1 dv2 (3:109) Àp Àp Example 3. jv2 ) ! then ax(n1 . jv2 ) given by & X( jv1 .29 Find the inverse DSFT of X( jv1 .v2 ) 2 D otherwise where D is the dashed area shown in Figure 3. n2 À b) eÀjav1 eÀjbv2 X( jv1 . .1 PROPERTIES 2-D DSFT Properties of 2-D DSFT are similar to the properties of 1-D DTFT. j(v2 þ 2p)] (c) Delay Property: If x(n1 . n2 ) ¼ 2 4p p p ð ð X( jv1 . n2 ) þ by(n1 .

SOLUTION Using inverse 2-D DSFT.Mathematical Foundations ω2 π D 133 –π π ω1 D –π FIGURE 3. n2 ) ¼ 0 > > (1Àcos : n1 ¼ n2 ¼ 0 n1 ¼ 0.12 Support of X(jv1 . n2 ) ¼ X( jv1 . jv2 ). n2 ) ¼ 1 1 À eÀjpn1 1 À eÀjpn2 1 ejpn1 À 1 ejpn2 À 1 þ 2 2 jn1 jn2 4p 4p jn1 jn2 which can be simpliﬁed to x(n1 . we have x(n1 . n2 6¼ 0 n1 6¼ 0. n2 ) ¼ Therefore. 8 > 0:5 > <0 x(n1 . n2 ¼ 0 otherwise 1 1 À eÀjpn1 1 À eÀjpn2 1 ejpn1 À 1 ejpn2 À 1 þ 2 jn1 jn2 4p2 4p jn1 jn2 n1 p)(1Àcos n2 p) 2p2 n1 n2 . jv2 )ejn1 v1 ejn2 v2 dv1 dv2 Àp Àp The above integral is 1 x(n1 . we have 1 4p2 p p ð ð x(n1 . n2 ) ¼ 4p2 0 0 ð ð e Àp Àp jn1 v1 jn2 v2 e 1 dv1 dv2 þ 2 4p p p ðð ejn1 v1 ejn2 v2 dv1 dv2 0 0 Upon integration.

det (lI À A) ¼ 0 (3:112) The above determinant is a polynomial of degree n and is denoted by P(l). that is.9. For example the steadystate response of an LTI system to an input of a sinusoidal signal is a sinusoidal signal with same frequency as that of the input.9 EIGENVALUES AND EIGENVECTORS Spectral analysis of matrices through use of eigenvalues=eigenvectors and SVD plays an important role in analysis and design of control systems using state-space approach. but different magnitude and phase.1 DEFINITION OF EIGENVALUE AND EIGENVECTOR The nonzero vector x is an eigenvector of square n Â n matrix A if there is a scale factor l such that Ax ¼ lx (3:110) The scale factor l is called the eigenvalue corresponding to the eigenvector x.30 Find the eigenvalues and eigenvectors of the 2 Â 2 matrix A: A¼ À2 2 À24 12 ! . The above equation can be considered as an operator operating on x. Solution of LTI continuous and discrete systems is also directly related to functions of matrices that is computed using eigenvalue and eigenvector decomposition. This means that eigenvectors are invariant with respect to operator A. Corresponding to each eigenvalue there is an eigenvector. The eigenvectors of A are vectors that are not changed by the operator. This polynomial is called the characteristic polynomial of matrix A. The eigenvalues can be repeated eigenvalues and also they may be complex. This is similar to the concept of eigenfunctions of LTI systems. The characteristic polynomial has n roots that are eigenvalues of matrix A.134 Control of Color Imaging Systems: Analysis and Design 3. Example 3. they are only scaled by l. The equation Ax ¼ lx can be written as (lI À A)x ¼ 0 (3:111) This equation has a nontrivial solution if and only if matrix lI À A is singular. Techniques such as state feedback by pole placement and design of state estimators are all based on eigenvalues=eigenvectors decomposition. sinusoidal signals are eigenfunctions of LTI systems. Therefore. 3.

Eigenvector corresponding to l1 ¼ 4 is computed as Ax1 ¼ l1 x1 8 ! ! ! < Àa þ b ¼ 2a a a 2 ! ¼4 b 12 b : À12a þ 6b ¼ 2b À2 À24 ! b ¼ 3a Let a ¼ 1. then b ¼ 3 and x1 ¼ ! ! 1 a ¼ 3 b Eigenvector corresponding to l2 ¼ 6 is given by Ax2 ¼ l2 x2 or À1 À12 1 6 ! 8 ! ! < Àa þ b ¼ 3a a a ! ! ¼3 b b : À12a þ 6b ¼ 3b b ¼ 4a Let a ¼ 1. The characteristic polynomial of matrix A is lþ 2 À2 ¼ l2 À 10l þ 24 ¼ (l À 4)(l À 6) ¼ 0 P(l) ¼ jlI À Aj ¼ 24 l À 12 Therefore. then b ¼ 4 and x2 ¼ ! ! 1 a ¼ 4 b Example 3. Now we compute the eigenvectors corresponding to the two eigenvalues.31 Find the eigenvalues and eigenvectors of the 3 Â 3 matrix A: 2:6 A ¼ 4 0:8 0:8 2 3 1:3 À2:5 5:4 À5 5 1:4 À1 SOLUTION The characteristic polynomial of matrix A is P(l) ¼ jlI À Aj ¼ l3 À 7l2 þ 14l À 8 ¼ (l À 1)(l À 2)(l À 4) ¼ 0 .Mathematical Foundations 135 SOLUTION First we compute the two eigenvalues l1 and l2 . l1 ¼ 4 and l2 ¼ 6.

136 Control of Color Imaging Systems: Analysis and Design The eigenvalues are l1 ¼ 1. we have & 1:3b À 2:5c ¼ À1:6 4:4b À 5c ¼ À0:8 Similarly. The corresponding eigenvectors are computed as 2:6 1:3 Ax1 ¼ l1 x1 ! 4 0:8 5:4 0:8 1:4 2 32 3 2 3 ( À2:5 a a 2:6a þ 1:3b À 2:5c ¼ a À5 54 b 5 ¼ 4 b 5 ! 0:8a þ 5:4b À 5c ¼ b À1 c c 0:8a þ 1:4b À c ¼ c These three equations are not linearly independent. ( 2:6a þ 1:3b À 2:5c ¼ 2a 0:8a þ 5:4b À 5c ¼ 2b 0:8a þ 1:4b À c ¼ 2c 8 < b ¼ 0:5 2 2 3 32 3 À2:5 a a 54 b 5 ¼ 24 b 5 À5 À1 c c 8 <b ¼ 4 3 :c ¼ 4 3 ! 2 3 3 16 7 ! x1 ¼ 4 4 5 3 4 2 & 1:3b À 2:5c ¼ À0:6 3:4b À 5c ¼ À0:8 ! 7 6 7 6 ! x2 ¼ 6 0:5 7 : 5 4 c ¼ 0:5 0:5 1 3 The third eigenvector is 2:6 1:3 Ax3 ¼ l3 x3 ! 4 0:8 5:4 0:8 1:4 Hence. and l3 ¼ 4. Only two of them are linearly independent. l2 ¼ 2. 2:6 1:3 Ax2 ¼ l2 x2 ! 4 0:8 5:4 0:8 1:4 Therefore. ( 2:6a þ 1:3b À 2:5c ¼ 4a 0:8a þ 5:4b À 5c ¼ 4b 0:8a þ 1:4b À c ¼ 4c ( ! b¼3 c¼1 2 3 1 6 7 ! x3 ¼ 4 3 5 1 2 32 3 2 3 À2:5 a a À5 54 b 5 ¼ 44 b 5 À1 c c & 1:3b À 2:5c ¼ À1:4 1:4b À 5c ¼ À0:8 . Using the ﬁrst two equations with a ¼ 1.

Expanding determinant of M11 . we have l1 l2 Á Á Á ln ¼ jAj Now we show that the sum of eigenvalues is equal to the trace 2 l À a11 Àa12 ÁÁÁ Àa1n 6 Àa21 l À a22 Á Á Á Àa2n 6 P(l) ¼ jlI À Aj ¼ 6 . Also from Equation 3. .9. then P(l) ¼ jlI À Aj ¼ (l À l1 )(l À l2 ) Á Á Á (l À ln ) Set l ¼ 0.. 4 . Àan1 Àan2 (3:116) of the matrix: 3 7 7 7 5 (3:117) (3:115) (3:114) (3:113) Á Á Á l À ann By expanding the above determinant along the ﬁrst column. . Therefore. Let A be an n Â n matrix with characteristic polynomial P(l). then jÀAj ¼ (Àl1 )(Àl2 ) Á Á Á (Àln ) ¼ (À1)n l1 l2 Á Á Á ln or (À1)n jAj ¼ (À1)n l1 l2 Á Á Á ln Hence.Mathematical Foundations 137 OF 3. we have 0 jM11 j ¼ (l À a22 )jM11 j À n X i¼3 0 (À1)iþ1 ai1 jMi1 j (3:119) Continuing these expansions.2 PRODUCT AND SUM EIGENVALUES The product and sum of eigenvalues of any matrix are equal to the determinant and the trace of that matrix. we have P(l) ¼ (l À a11 )jM11 j À n X i¼2 (À1)iþ1 ai1 jMi1 j (3:118) where Mij is the determinant of matrix obtained by deleting the ith row and the jth column of A. l1 þ l2 þ Á Á Á þ ln ¼ a11 þ a22 þ Á Á Á þ ann ¼ Trace(A) (3:121) . . . Therefore. the second leading coefﬁcient of P(l) is l1 þ l2 þ Á Á Á þ ln . . We ﬁrst show that the product of eigenvalues is equal to the determinant of the matrix. . respectively. .113. we will have P(l) ¼ (l À a11 )(l À a22 ) Á Á Á (l À ann ) þ P0 (l) (3:120) where P0 (l) is a polynomial of degree n À 2. . the leading coefﬁcient of P(l) is one and the second leading coefﬁcient is a11 þ a22 þ Á Á Á þ ann . .

The characteristic polynomial of matrix A can be found by using the following recursive algorithm. . n. Example 3.138 Control of Color Imaging Systems: Analysis and Design OF A 3. k ¼ 1. .3 FINDING CHARACTERISTIC POLYNOMIAL MATRIX Let A be an n Â n matrix. . A2 .9. . . . 1 an ¼ À (anÀ1 W1 þ anÀ2 W2 þ Á Á Á þ a1 WnÀ1 þ Wn ) n and P(l) ¼ ln þ a1 lnÀ1 þ a2 lnÀ2 þ Á Á Á þ anÀ1 l þ an The above algorithm is known as Bocher’s formula [4].32 Find the characteristic polynomial of the following 3 Â 3 matrix: À2 A¼4 1 1 2 3 4 2 6 3 5 À1 À5 (3:122) (3:123) SOLUTION The trace of A. and A3 are 2 6 A2 ¼ A Â A ¼ 4 1 2 1 10 À2 4 6 14 37 3 76 3 54 1 À2 24 32 1 À2 1 2 32 À2 4 6 4 6 À1 7 6 3 5¼4 7 2 3 2 3 2 10 14 37 3 126 245 À38 À2 24 72 3 7 5 5 3 À1 À5 À1 À5 6 A3 ¼ A2 Â A ¼ 4 7 76 5 54 1 À8 2 À8 7 6 3 5 ¼ 4 28 43 7 100 5 À127 À8 W1 ¼ Trace(A) ¼ À1 W2 ¼ Trace(A2 ) ¼ 71 W3 ¼ Trace(A3 ) ¼ 110 À5 . then the coefﬁcients of the characteristic equation are [3] a1 ¼ ÀW1 a2 ¼ À 1 (a1 W1 þ W2 ) 2 a3 ¼ À 1 (a2 W1 þ a1 W2 þ W3 ) 3 . 2. Let Wk ¼ Trace(Ak ).

the modal matrix is 7 6 x2 ¼ 4 0:5 5.Mathematical Foundations Hence. a1 ¼ ÀW1 ¼ 1 a2 ¼ À 1 (a1 W1 þ W2 ) ¼ À 1 (À1 Â 1 þ 71) ¼ À35 2 2 a3 ¼ À 1 (a2 W1 þ a1 W2 þ W3 ) ¼ À 1 (35 Â 1 þ 1 Â 71 þ 110) ¼ À72 3 3 Therefore.33 Find the modal matrix of 2 6 A ¼ 4 0:8 0:8 2:6 1:3 À2:5 5:4 1:4 3 7 À5 5 À1 SOLUTION Matrix A has three independent eigenvectors (see Example 3. 0:5 2 M ¼ ½ x1 x2 ÁÁÁ 6 xn ¼ 4 1:333 1:333 1 7 0:5 3 5 0:5 1 1 1 3 . The n Â n matrix M formed by side-by-side stacking of eigenvectors is called the modal matrix of A: M ¼ ½ x1 x2 Á Á Á xn (3:124) Since the n columns of M are linearly independent. P(l) ¼ l3 þ 8l2 þ 9l À 2 139 3. . Example 3. .9.31): 2 1 3 2 1 3 and 2 3 1 6 7 x3 ¼ 4 3 5 1 7 6 x1 ¼ 4 1:3333 5. . . it is full rank and hence invertible. xn be the n independent eigenvectors of n Â n matrix A. 1:3333 Therefore. x2 .4 MODAL MATRIX Let x1 .

it is diagonalizable if and only if the eigenvectors are linearly independent. x2 . l2 . if the eigenvectors are linearly independent.. .5 MATRIX DIAGONALIZATION Let A be an n Â n matrix with n distinct eigenvalues l1 .9. . .140 Control of Color Imaging Systems: Analysis and Design 3. Therefore. 0 ÁÁÁ ÁÁÁ . Axn ¼ ln xn These equations can be put together in matrix form to obtain ½ Ax1 Ax2 ÁÁÁ Axn ¼ ½ l1 x1 l 2 x2 ÁÁÁ ln x n (3:128) (3:127) (3:125) (3:126) Equation 3. then the above matrix equation can be written in terms of L and modal matrix M as AM ¼ ML (3:130) This equation is valid regardless of whether the eigenvectors are linearly independent or not. . . ln . Assume that the corresponding independent eigenvectors are x1 . 0 2 0 l2 . . . . 4 . . . xn . However. we can post-multiply the above equation by M À1 to obtain A ¼ MLM À1 or L ¼ M À1 AM (3:132) (3:131) If matrix A has repeated eigenvalues. . then M is full rank and has an inverse and in this case. . . 0 0 . 3 7 7 7 5 (3:129) A½ x1 x2 ÁÁÁ xn ¼ ½ x1 x2 Á Á Á ln Deﬁne L to be the diagonal matrix of eigenvalues. we have Ax1 ¼ l1 x1 Ax2 ¼ l2 x2 .128 can be written as l1 60 6 Á Á Á xn 6 . . . . .

35 Diagonalize matrix A: 5 A¼4 2 À1 2 3 1 0 4 05 1 6 SOLUTION The eigenvalues and eigenvectors of A are l1 ¼ 3 3 and l2 ¼ l3 ¼ 6 2 3 2 3 0 23 x2 ¼ 4 0 5. l3 ¼ 4. 1 6 MÀ1 AM ¼ 4 1:333 À0:6 6 ¼ 4 1:6 0 2 1 0 6 ¼ 40 2 0 0 2 1:333 2 1 0:5 0:5 À0:3 À0:2 0:5 3 0 7 05 4 3À1 2 2:6 1:3 1 7 6 3 5 4 0:8 5:4 1 32 1 À2:5 76 À5 54 1:333 1 0:5 3 1 7 35 1:333 0:5 1 0:8 1:4 À1 32 32 3 1 1 1 2:6 1:3 À2:5 1:5 76 76 7 À1 54 0:8 5:4 À5 54 1:333 0:5 3 5 0:8 1:4 À1 À0:5 1:333 0:5 1 Example 3.34 Diagonalize 3 Â 3 matrix A given by 2:6 A ¼ 4 0:8 0:8 2 3 1:3 À2:5 5:4 À5 5 1:4 À1 SOLUTION The eigenvalues and modal matrix of A are 1 M ¼ 4 1:333 1:333 2 1 0:5 0:5 3 1 35 1 l1 ¼ 1. l2 ¼ 2. 1 2 . and Therefore.Mathematical Foundations 141 Example 3. and x3 ¼ 4 23 5 1 18 1 x1 ¼ 4 À2 5.

6 DEFINITE MATRICES Positive and negative deﬁnite (semi-deﬁnite) matrices are an important class of matrices with applications in signal processing. the quantity xH Ax 0. therefore. The following theorem states the condition for positiveness of a Hermitian matrix. and control systems. À2 0 Ax ¼ lx ! À4 1 À4 ! 3. Example 3. We now deﬁne four types of deﬁnite matrices: (a) Positive Deﬁnite Matrices: The n Â n Hermitian matrix A is said to be positive deﬁnite if for any nonzero vector x 2 Rn .9. the quantity xH Ax > 0.142 Control of Color Imaging Systems: Analysis and Design The repeated eigenvalues have independent eigenvectors. Matrix A is said to be Hermitian if A ¼ AH . (b) Positive Semi-Deﬁnite Matrices: The n Â n Hermitian matrix A is said to be positive semi-deﬁnite if for any nonzero vector x 2 Rn .36 Diagonalize the 2 Â 2 matrix A: A¼ 0 À4 1 À4 ! SOLUTION The eigenvalues of A are l1 ¼ l2 ¼ À2. They are particularly useful matrices in optimization problems. there is one eigenvector x1 ¼ and the matrix A is not diagonalizable. (d) Negative Semi-Deﬁnite Matrices: The n Â n Hermitian matrix A is said to be negative semi-deﬁnite if for any nonzero vector x 2 Rn . (c) Negative Deﬁnite Matrices: The n Â n Hermitian matrix A is said to be negative deﬁnite if for any nonzero vector x 2 Rn . Hence. Here H stands for conjugate transpose. matrix A is diagonalizable. The eigenvectors of A are computed as ( ! ! b ¼ À2a a a ! b ¼ À2a ! ¼ À2 b b À4a À 4b ¼ À2b ! 1 Therefore. the quantity xH Ax < 0. image processing. the quantity xH Ax ! 0. . 3À1 2 5 1 0 23 4 À2 0 23 5 4 2 M AM ¼ À1 1 1 18 2 À1 32 1 0 1 0 4 0 54 À2 0 1 6 1 1 3 2 23 3 23 5 ¼ 4 0 18 0 3 0 0 6 05 0 6 Note that if matrix A has repeated eigenvalues. it is not always possible to diagonalize it as is shown in the following example.

Similar theorems can be stated for positive semi-deﬁnite. then all of its eigenvalues are positive. Here we state a summary of the results of these theorems: (a) (b) (c) (d) Positive deﬁnite: All eigenvalues are positive. then Equation 3. Negative deﬁnite: All eigenvalues are negative. negative deﬁnite.1 An n Â n Hermitian matrix is positive deﬁnite if and only if all its eigenvalues are positive. li > 0. Let xi be the eigenvector corresponding to the ith eigenvalue li of A.Mathematical Foundations 143 THEOREM 3. Proof: First assume that A is Hermitian and all of its eigenvalues are positive.37 Check the following symmetric matrices for their deﬁniteness: À2 5 ! 2 1 (b) B ¼ 1 À2 (a) A ¼ 5 À2 (c) C ¼ À6 À4 À4 À6 ! ! . then xH Ax ¼ xH MLM H x Let y ¼ M H x.133 can be written as xH Ax ¼ xH MLM H x ¼ yH Ly ¼ n X i¼1 (3:133) li jyi j2 > 0 (3:134) Next we need to show that if A is Hermitian and positive deﬁnite. and negative semi-deﬁnite matrices. (3:135) Example 3. Negative semi-deﬁnite: All eigenvalues are nonpositive (zero or negative). Positive semi-deﬁnite: All eigenvalues are nonnegative (zero or positive). then 0 < xH Axi ¼ xH li xi ¼ li kxi k2 i i Therefore.

vn . Both are positive. sr ). . it has nonnegative eigenvalues. and ﬁnally. Let V1 ¼ ½ v1 v2 Á Á Á vr . then there exist unitary matrices U(m Â m) and V(n Â n) such that A ¼ USV H where S is an m Â n matrix with entries & Sij ¼ si 0 if i ¼ j if i 6¼ j (3:137) (3:136) The quantities s1 ! s2 ! Á Á Á ! sr > srþ1 ¼ srþ2 ¼ Á Á Á ¼ sn ¼ 0 are called singular values of A. the eigenvalues of matrix C are À2 and À10 that makes it negative deﬁnite. Notice that U1 is a unitary matrix of size m Â r. then AH AV1 ¼ V1 L2 (3:138) H Pre-multiply both sides of Equation 3. The theory for general matrices was established by Eckart and Young. v2 .139 we have U1 U1 ¼ I.10 SINGULAR VALUE DECOMPOSITION One of the most important tools in signal processing and numerical linear algebra is the SVD.2361 and 2. Let the eigenvalues and eigenvectors of S be s2 ! s2 ! Á Á Á ! s2 > s2 ¼ s2 ¼ Á Á Á ¼ s2 ¼ 0 1 2 r rþ1 rþ2 n and v1 . V2 ¼ ½ vrþ1 vrþ2 Á Á Á vn .144 Control of Color Imaging Systems: Analysis and Design SOLUTION The eigenvalues of A are 3 and 7. We ﬁrst state the SVD theorem and then we look at its applications.and pre-multiÀ1 plication by L results in H LÀ1 V1 AH AV1 LÀ1 ¼ I (3:139) H Choose U1 ¼ AV1 LÀ1 . s2 . Therefore. Eigenvalues of matrix B are À2. Choose U2 to be another unitary matrix of size m Â (m À r) orthogonal to U1 . hence matrix A is positive deﬁnite matrix. . These eigenvectors form an orthonormal set. Then . The SVD was discovered for square matrices by Beltrami and Jordan in the eighteenth century. .2361. therefore B is indeﬁnite. . THEOREM 3. and L ¼ diag(s1 . then by Equation 3. . . . . Proof: Let S ¼ AH A. 3.2 Let A be an m Â n real or complex matrix with rank r. Matrix S is n Â n Hermitian and positive semi-deﬁnite with rank r.138 by V1 followed by post.

The corresponding 1 2 3 eigenvectors are 3 À0:3651 v1 ¼ 4 0:1826 5. and s2 ¼ 0. s2 ¼ 52.Mathematical Foundations 145 " U AV ¼ ½ U1 U2 A½ V1 V2 ¼ H H H H U1 AV1 U1 AV2 H H U2 AV1 U2 AV2 # ¼ " L 0 0 0 # ¼ S (3:140) Pre.140 by U and V H . À0:9129 Therefore. 0:1961 2 3 À0:9309 v3 ¼ 4 À0:0716 5 0:3581 2 and and U1 ¼ AV1 LÀ1 ¼ À0:707 0:707 . 3 À0:3651 0 V1 ¼ 4 0:1826 À0:9806 5 À0:9129 0:1961 and 2 4 ¼ À2 6 À0:707 À0:707 ! 6 À4 ! 2 À0:3651 6 4 0:1826 À0:9129 3 " pﬃﬃﬃﬃﬃﬃ #À1 0 60 0 7 À0:9806 5 pﬃﬃﬃﬃﬃﬃ 0 52 0:1961 2 3 À0:9309 V2 ¼ 4 À0:0716 5 0:3581 2 2 3 0 v2 ¼ 4 À0:9806 5. Example 3.38 Find the SVD of matrix A given by A¼ 2 À2 4 6 6 À4 ! SOLUTION We ﬁrst form S ¼ AT A 8 À4 S ¼ A A ¼ 4 À4 52 20 0 T 2 3 20 0 5 52 The eigenvalues of S are s2 ¼ 60. we have A ¼ USV H (3:141) The nonnegative numbers s1 ! s2 ! Á Á Á ! sr > srþ1 ¼ srþ2 ¼ Á Á Á ¼ sn ¼ 0 are called singular values of A and they are square roots of eigenvalues of AH A or AAH .and post-multiplying both sides of Equation 3.

4 0. . most of the singular values are small. each component is an image of size 256 Â 256.5 0.3 0. the p matrix norm is deﬁned by kAkp ¼ sup kAxkp x2Rn kxkp ¼1 (3:142) 1 0.9 0.1 0 0 50 100 150 k 200 250 300 FIGURE 3. By deﬁnition. Therefore.8 0. respectively.146 Control of Color Imaging Systems: Analysis and Design Since m ¼ 2 and r ¼ 2. The ith component of the image is s2 ui vi and the whole image is i A ¼ USV T ¼ 256 X i¼1 si ui viT Plot of the 256 singular values of the image normalized with respect to the largest singular value is shown in Figure 3. we have U ¼ U1 .39 Consider the 256 Â 256 cameraman image. we have À0:707 À0:707 A ¼ USV ¼ 0:707 À0:707 H 2 3T ! pﬃﬃﬃﬃﬃﬃ ! À0:3651 0 À0:9309 60 p0 ﬃ 0 4 ﬃﬃﬃﬃﬃ 0:1826 À0:9806 À0:0716 5 0 52 0 À0:9129 0:1961 0:3581 Example 3.13 Singular values normalized with respect to the largest singular value. As can be seen.1 MATRIX NORM Matrix norm like other vector-space norms must satisfy the properties of vector norm.2 0.10. 3.16.7 σ 2(k)/σ 2 max 0.6 0.13. If we decompose the image into its SVD components. The original image and the images reconstructed using the ﬁrst 20 and 50 singular values are shown in Figures 3.14 through 3. Let A be an m Â n matrix mapping vector space Rn to vector space Rm .

FIGURE 3.Mathematical Foundations 147 FIGURE 3. .14 Cameraman image.15 Image reconstructed using the ﬁrst 10 eigenimages. FIGURE 3.16 Image reconstructed using the ﬁrst 50 eigenimages.

148 Control of Color Imaging Systems: Analysis and Design ||Ax|| x2 y = Ax y2 x1 x= x1 x2 y = y1 y1 y2 ||x|| = 1 FIGURE 3.17 Matrix norm. and p ¼ 1. if the vector norm is l1 norm then the matrix norm is based on vector l1 norm. We consider three cases corresponding to p ¼ 1. The matrix norm depends on vector norm. The concept of matrix norm for a 2 Â 2 matrix is illustrated in Figure 3. kAk1 is equal to the longest row sum. which means that to ﬁnd the p ¼ 1 norm. the matrix norm becomes k Ak2 ¼ max kAxk2 kxk2 ¼1 (3:145) . for example. compute the sum of absolute values of each row and pick the maximum. Matrix norm is a measure of boundness of that matrix. compute the sum of absolute values of each column and pick the maximum. kAk1 is equal to the longest column sum. p ¼ 2. Case III: p ¼ 2 In this case. the matrix norm becomes kAk1 ¼ max kAxk1 ¼ max kxk1 ¼1 i m X j¼1 jaij j (3:144) Therefore. Case I: p ¼ 1 In this case.17. the matrix norm becomes kAk1 ¼ max kAxk1 ¼ max kxk1 ¼1 j m X aij i¼1 (3:143) Therefore. which means that to ﬁnd the p ¼ 1 norm. Case II: p ¼ 1 In this case.

kAk2 ¼ max l(AH A) (3:150) The p norm has the property that for any two matrices A and B. It is deﬁned as kAkF ¼ m n X X 2 aij i¼1 j¼1 !1 2 (3:153) Frobenius norm is also called Euclidean norm. we must solve the following optimization problem: maximize kAxk2 ¼ xH AH Ax x subject to: xH x ¼ 1 (3:146) Using Lagrange multiplier technique. the following inequality holds: kAxkp and kABkp kAkp kBkp (3:152) kAkp kxkp (3:151) Frobenius Norm Frobenius norm is another matrix norm that is not a p norm. ﬃﬃﬃ a simple example. The Frobenius norm can also be expressed as pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ (3:154) kAkF ¼ Trace(AH A) . l must be chosen to be the maximum eigenvalue of the positive deﬁnite matrix AH A.Mathematical Foundations 149 To ﬁnd the p ¼ 2 matrix norm. therefore. the solution for vector x must be an eigenvector of square matrix AH A corresponding to eigenvalue l and the resulting norm is kAxk2 ¼ xH AH Ax ¼ lxH x ¼ l (3:149) Since we are maximizing the norm. this is equivalent to the following optimization problem: maximize J ¼ xH AH Ax À l(xH x À 1) x (3:147) Setting the gradient of J with respect to x equal to zero. the Frobenius p As norm of an n Â n identity matrix is kIkF ¼ n. we obtain the equation qJ ¼ 2AH Ax À 2lx ¼ 0 ! AH Ax ¼ lx qx (3:148) Therefore.

. p ¼ 2.10.150 Control of Color Imaging Systems: Analysis and Design Example 3. and let X1 . X2 . XN be N observations of random vector X. The PCA is extremely useful in case of data of higher dimensions since illustrative techniques are not applicable in case of data of dimension greater than three. . similarities.40 Find the p ¼ 1. . . p ¼ 1. and differences hidden in data. and Frobenius norm of the following matrix: A¼ 4 2 À6 7 ! SOLUTION The different matrix norms are computed as kAk1 ¼ max j m X aij ¼ max i¼1 m X j¼1 j m P i¼1 jai1 j a1j m P i¼1 ! jai2 j ¼ max ½ 10 j 9 ¼ 10 kAk1 ¼ max i aij ¼ max i m P j¼1 m P j¼1 ! a2j ¼ max ½ 6 13 ¼ 13 i pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ kAk2 ¼ max l(AH A) ¼ sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ! ! 4 À6 4 2 max l 2 7 À6 7 sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ! pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ 52 À34 ¼ max l ¼ 86:5037 ¼ 9:3007 À34 53 0 11 2 2 2 X X 2 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ pﬃﬃﬃﬃﬃﬃﬃﬃ aij A ¼ 16 þ 4 þ 36 þ 49 ¼ 105 ¼ 10:247 kAkF ¼ @ i¼1 j¼1 3. An important application of PCA is data dimensionality reduction or data compression.2 PRINCIPAL COMPONENTS ANALYSIS Principal component analysis (PCA) is a statistical technique to extract patterns. Let X be an m-dimensional random vector. This is achieved by reducing the data dimension with a small loss of information. The m Â m covariance matrix R of these data is estimated using the sample covariance by R¼ N 1 X (Xn À m)(Xn À m)T N n¼1 (3:155) where m is the sample mean vector given by m¼ N 1 X Xn N n¼1 (3:156) .

qj ¼ N X i¼1 aij Xi ¼ aT X j j ¼ 1. . . ^ X¼ where ai ¼ X T qi Example 3. . . j) of the LENA image.41 Figure 3. that is. Once the principal components are established. that is. . qr are the normalized eigenvectors corresponding to the r largest eigenvalues of R. . Therefore. . . 2. . The solution is j j the normalized eigenvector corresponding to the largest eigenvalue of the sample covariance matrix R.18 shows 2-D scatter data obtained by plotting two neighboring pixels x ¼ f (i. . . The estimated covariance matrix of this data set is R¼ 0:2529 0:2512 0:2512 0:2525 ! r X i¼1 ai q i (3:159) i ¼ 1. the r principal components q1 . . . 2. r (3:157) where T aj ¼ ½ a1j a2j Á Á Á aNj is a unit norm vector X ¼ ½ X1 X2 Á Á Á XN is a data matrix The sample variance of qj is given by s2 ¼ j T 1 T 1 qj À aT m ¼ aT X À aT m qi À aT m a T X À aT m j j j j j N N j N 1 1 X (Xi À m)T (Xi À m)aj (X À m)T (X À m)aj ¼ aT j N N i¼1 ¼ aT j ¼ aT Raj j (3:158) We now maximize s2 ¼ aT Raj with respect to unit norm vector aj . . X2 . XN . . q2 . any random vector in this class can be approximated by linear combinations of these principal components. j) and y ¼ f (i þ 1. r (3:160) . .Mathematical Foundations 151 Let q1 . q2 . Note that the principal components are pairwise orthogonal. . qr be r vectors of size m Â 1 obtained as linear combinations of X1 . . . .

q2 ¼ À0:7074 l2 ¼ 0:0014 ! 0:7074 .1 Control of Color Imaging Systems: Analysis and Design 0.152 0. To compress the image. These eigenimages are used as basis functions for compressing the image. the basis functions.6 0.18. and the error image are shown in Figures 3.19 through 3.3 0.42 In this example.4 0. Note that the basis functions are not quantized. The two eigenvectors and the corresponding eigenvalues of covariance matrix R are q1 ¼ and ! 0:7068 .22.5 0. Example 3. The image is partitioned into 8 Â 8 block and the 64 Â 64 sample covariance matrix is formed.1 0.2 0. The original image. the compressed image. we use 16 of these principal component images corresponding to the 16 largest singular values.8 0.3 0.9 FIGURE 3.7 0.6 0.4 0. The peak signal to noise ratio (PSNR) of the compressed image is 65 dB.8 0.7 0.5 0. If we quantize the basis . 0:7078 l1 ¼ 0:5039 The two principal components are also shown in Figure 3. There are 64 basis functions.18 LENA image scatter data with the two principal components. the gray-scale cameraman image is used to obtain the principal components (eigenimages). respectively.2 0. Each basis function is an 8 Â 8 image.9 0.

20 Eigenimages.21 Compressed image.Mathematical Foundations 153 FIGURE 3. . FIGURE 3.19 Original image. FIGURE 3.

the compressed image. respectively. The PSNR of the FIGURE 3.43 To compress an RGB image. the PSNR will decrease and the compression loss will be more visible. we transform the image from RGB color space to YCb Cr color space [5]. To compress the image.) Original RGB image. Each basis function is an 8 Â 8 image.25.22 Error image. . functions and the weights of the basis function. The original image. There are 64 basis functions per channel. and the error image are shown in Figures 3.154 Control of Color Imaging Systems: Analysis and Design FIGURE 3.23 (See color insert following page 428. we use 8 of these principal component images corresponding to the 8 largest singular values.23 through 3. The three channels are independently partitioned into 8 Â 8 blocks and the 64 Â 64 sample covariance matrix is estimated for each channel. Example 3.

25 (See color insert following page 428. They are useful for high-quality color imaging systems.11 MATRIX POLYNOMIALS AND FUNCTIONS OF SQUARE MATRICES Matrix polynomials and functions of matrices have application in different engineering discipline such as modern control theory. FIGURE 3.) Error RGB image. solution of simultaneous linear DEs.) Compressed RGB image. Again no quantization is performed on the basis of functions and weights. 3.24 (See color insert following page 428. and communication systems.Mathematical Foundations 155 FIGURE 3. In this section. queuing theory. compressed image is 55 dB. . we focus on matrix polynomials and some special function of square matrices with their analytical computations.

. where li is the ith eigenvalue of matrix A. . a2 . The matrix polynomial f (A) can be factored as f (A) ¼ (A À a1 I)(A À a2 I) Á Á Á (A À an I) (3:164) 3. For example.1 MATRIX POLYNOMIAL Consider the monic polynomial f (x) of degree m given by f (x) ¼ xm þ a1 xmÀ1 þ a2 xmÀ2 þ Á Á Á þ amÀ1 x þ am (3:161) If the scalar variable x is replaced by the n Â n matrix A.11. . the geometric matrix series S(A) ¼ I þ aA þ a2 A2 þ Á Á Á ¼ is a convergence series and converges to S(A) ¼ (I À aA)À1 If and only if the scalar geometric series S(l) ¼ 1 þ al þ a2 l2 þ Á Á Á ¼ 1 X k¼0 1 X k¼0 ak Ak (3:166) (3:167) (al)k (3:168) .11.156 Control of Color Imaging Systems: Analysis and Design 3.2 INFINITE SERIES OF MATRICES 1 X k¼0 An inﬁnite series of matrix A is deﬁned as S(A) ¼ a0 I þ a1 A þ a2 A2 þ Á Á Á ¼ ak A k (3:165) It can be shown that the matrix inﬁnite series S(A) converges if and only if the scalar inﬁnite series S(li ) converges for all values of i. an are the roots of the polynomial f (x). then the corresponding matrix polynomial f (A) is deﬁned by f (A) ¼ Am þ a1 AmÀ1 þ a2 AmÀ2 þ Á Á Á þ amÀ1 A þ am I where Am ¼ A Â m times (3:162) A Â ÁÁÁ Â A I is an n Â n identity matrix The polynomial f (x) can be written in factor form f (x) ¼ (x À a1 )(x À a2 ) Á Á Á (x À an ) (3:163) where a1 . .

Mathematical Foundations 157 converges for all eigenvalues of A. .3 Any n Â n square matrix satisﬁes its own characteristic polynomial. P(A) ¼ n X i¼0 n X i¼0 (3:171) (3:172) ai AnÀi (3:173) ai AnÀi ¼ n X i¼0 ai MLnÀi M À1 ¼ M 0 0 . 6 . 4 .11. .3 CAYLEY–HAMILTON THEOREM THEOREM 3. Then P(l) is a polynomial of degree n in l and is given by P(l) ¼ ln þ a1 lnÀ1 þ a2 lnÀ2 þ Á Á Á þ anÀ1 l þ an Then P(A) ¼ An þ a1 AnÀ1 þ a2 AnÀ2 þ Á Á Á þ anÀ1 A þ an I ¼ Since Ak ¼ MLk M À1 . . . we must have jali j < 1 or À 1 1 < li < jaj jaj for i ¼ 1. With the assumption that eigenvalues of A are distinct. n (3:169) Another example is the exponential matrix polynomial deﬁned by eA ¼ I þ A þ 1 X Ak A2 A3 þ þ ÁÁÁ ¼ 2! 3! k! k¼0 (3:170) 3. . 2. .. Proof: We prove the theorem for a special case when matrix A has n distinct eigenvalues. . that is. therefore. . 0 2 0 ÁÁÁ P(l2 ) Á Á Á . . . P(A) ¼ 0 where P(l) ¼ jlI À Aj is the characteristic polynomial of matrix A. P(ln ) 3 n X i¼0 ai LnÀi M À1 P(l1 ) 6 0 6 ¼ M6 . Let P(l) be the characteristic polynomial of A. 0 ÁÁÁ 7 7 À1 7M 7 5 (3:174) .

4 Consider an n Â n matrix A and a matrix polynomial in A of degree m ! n. This is proved in the following theorem. THEOREM 3.44 Show that matrix A satisﬁes its characteristic polynomial: A¼ 2 5 À7 1 ! (3:175) SOLUTION The characteristic polynomial of A is lÀ2 P(l) ¼ jlI À Aj ¼ À5 Then P(A) ¼ A2 À 3A þ 37I ! ! 2 À7 2 À7 ¼ À3 5 1 5 1 ! À31 À21 À6 ¼ þ 15 À34 À15 ! ! 2 À7 1 0 þ 37 5 1 0 0 ! ! ! 21 37 0 0 0 þ ¼ ¼0 À3 0 37 0 0 7 ¼ l2 À 3l þ 37 lÀ 1 Cayley–Hamilton theorem is used to reduce any matrix polynomial of n Â n matrix A to a polynomial of degree n À 1. f (A) ¼ am Am þ amÀ1 AmÀ1 þ amÀ2 AmÀ2 þ Á Á Á þ a1 A þ a0 I Then f (A) is reducible to a polynomial of degree n À 1 given by R(A) ¼ rnÀ1 AnÀ1 þ rnÀ2 AnÀ2 þ rnÀ3 AnÀ3 þ Á Á Á þ r1 A þ r0 I (3:177) (3:176) Proof: Consider the scalar polynomial f (l). then f (l) ¼ Q(l)P(l) þ R(l) (3:178) . then P(A) ¼ 0 Example 3. That is. the characteristic polynomial of matrix A. Divide this polynomial by P(l).158 Control of Color Imaging Systems: Analysis and Design Since P(l1 ) ¼ P(l2 ) ¼ Á Á Á ¼ P(ln ) ¼ 0.

1 Cayley–Hamilton Technique First we assume that the eigenvalues of matrix A are distinct. respectively. 3.11. we have f (A) ¼ R(A) ¼ rnÀ1 AnÀ1 þ rnÀ2 AnÀ2 þ rnÀ3 AmÀ3 þ Á Á Á þ r1 A þ r0 I (3:180) . The degree of the remainder polynomial is at most n À 1 since the degree of characteristic polynomial is n. then we have f (A) ¼ Q(A)P(A) þ R(A) ¼ R(A) Example 3.4. Since P(A) ¼ 0 by the Cayley–Hamilton theorem. Using the results of the Cayley–Hamilton theorem. ! ! 4 À2 1 0 f (A) ¼ R(A) ¼ À239A þ 2041I ¼ À239 þ 2041 3 0 0 1 ! ! ! À956 478 2041 0 1085 478 ¼ þ ¼ À717 0 0 2041 À71 2041 3. They include Cayley– Hamilton and matrix diagonalization techniques.4 FUNCTION OF MATRICES There are several techniques to ﬁnd function of matrices. we have l8 À 8l6 þ 2l5 þ 2l3 À 3l2 þ l þ 7 ¼ l6 þ 4l5 þ 2l4 À 14l3 À 68l2 À 186l l2 À 4l þ 6 À239l þ 2041 À 339 þ 2 l À 4l þ 6 Therefore.45 Find the matrix polynomial f (A) ¼ A8 À 8A6 þ 2A5 þ 2A3 À 3A2 þ A þ 7I if A¼ 4 2 À7 0 ! (3:179) SOLUTION The characteristic polynomial of A is l À 4 2 2 P(l) ¼ jlI À Aj ¼ À3 l ¼ l À 4l þ 6 Dividing f (l) by P(l).11.Mathematical Foundations 159 where Q(l) and R(l) are the quotient and remainder polynomial after dividing f (l) by P(l).

. . . 6. we have the following n equations for n unknown r0 . . . 76 . . l1 . 1 1 l1 l2 l3 . . ln |ﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄ{zﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄ} mþ1 mþ2 m (3:184) Since l1 is an eigenvalue with multiplicity of m. ¼ . . f (ln ) ¼ R(ln ) (3:186) . . n in the above equation: f (li ) ¼ Q(li )P(li ) þ R(li ) ¼ R(li ) (3:182) This yield a set of n simultaneous linear equations that can be solved for r0 . ¼ . Without loss of generality. l . . l2 n ÁÁÁ ÁÁÁ ÁÁÁ . . 7 .160 Control of Color Imaging Systems: Analysis and Design where f (l) ¼ Q(l)P(l) þ R(l) (3:181) To get the coefﬁcients of the polynomial R(l). 2. r1 . . . . 7 6 . . rnÀ1 : 2 6 61 6 6 61 6 6. . then dk Q(l)P(l) ¼0 dlk l1 for k ¼ 1. That is. . ÁÁÁ lnÀ1 1 32 76 lnÀ1 76 2 76 76 lnÀ1 76 3 76 . . . . ln l2 1 l2 2 l2 3 . 5 f (ln ) f (l1 ) 3 (3:183) rnÀ1 Now let us assume that we have repeated eigenvalues. 7 . . 7 6 . . i ¼ 1. 5 4 . . . . . . . . . m (3:185) Therefore. 2. . 76 . . l 1 . . . r1 . l . rnÀ1 : f (l1 ) ¼ R(l1 ) f 0 (l1 ) ¼ R0 (l1 ) f 00 (l1 ) ¼ R00 (l1 ) . we assume that the n Â n matrix A has one eigenvalue of multiplicity m and n À m distinct eigenvalues. . . f mÀ1 (l1 ) ¼ RmÀ1 (l1 ) f (lmþ1 ) ¼ R(lmþ1 ) f (lmþ2 ) ¼ R(lmþ2 ) . we set l ¼ li . . . 54 lnÀ1 n r0 3 2 7 7 6 r1 7 6 f (l2 ) 7 7 7 6 7 7 6 r2 7 ¼ 6 f (l3 ) 7 7 7 6 . 4. the eigenvalues of A are l 1 .

then A ¼ MLM À1 (3:187) .2 Modal-Matrix Technique Assume that matrix A has distinct eigenvalues and let M be the modal matrix of A. we have r0 r1 Therefore.46 Find f (A) ¼ eAt if 0 A¼ À2 1 À3 ! SOLUTION The characteristic polynomial of A is jlI À Aj ¼ Hence the eigenvalues are l1 ¼ À1 and Since they are simple eigenvalues. eAt ¼ r0 I þ r1 A ¼ Substituting for r0 and r1 eAt ¼ r0 À2r1 r1 r0 À 3r1 ! ¼ 2eÀt À eÀ2t 2eÀt À eÀ2t 2eÀt À eÀ2t 2eÀt À eÀ2t ! r0 0 ! 0 0 þ r0 À2r1 r1 À3r1 ! ¼ r0 À2r1 r1 r0 À 3r1 ! ! ¼ 1 1 À1 À2 !À1 eÀt eÀ2t ! ¼ 2 À1 1 À1 ! eÀt eÀ2t ! ¼ 2eÀt À eÀ2t eÀt À eÀ2t ! À1 À2 ! r0 r1 ! ¼ eÀt eÀ2t ! l2 ¼ À2 l 2 ! À1 ¼ (l þ 1)(l þ 2) ¼ 0 lþ3 3.4.Mathematical Foundations 161 Example 3.11. we have f (l1 ) ¼ R(l1 ) ¼ r0 þ b1 l1 f (l2 ) ¼ R(l2 ) ¼ r1 þ b1 l2 or in matrix form 1 1 Solving for r0 and r1 .

. . . . 7 . f (A) ¼ M 6 . 0 0 . . 7 . ak L k ¼ 6 . 7 7 7 À1 7M 7 0 5 f (ln ) 0 0 . . .162 Control of Color Imaging Systems: Analysis and Design Let the Taylor series expansion of the function f (A) be f (A) ¼ 1 X k¼0 ak Ak (3:188) Now substituting Equation 3. . . . . ¼6 . . . . . .187 into Equation 3. 7 . . 7 . 6 6 7 0 Á Á Á f (lnÀ1 ) 0 5 4 0 0 Therefore.188 results in f (A) ¼ Since 0 ÁÁÁ 0 6 k¼0 6 6 1 P 6 0 ak lk Á Á Á 0 6 2 6 k¼0 1 6 X 6 . . . .. . 6 . f (l1 ) 0 6 0 f (l2 ) 6 6 . . 6 k¼0 6 1 P 6 0 0 ÁÁÁ ak lk 6 nÀ1 6 k¼0 6 4 0 0 ÁÁÁ 0 2 3 f (l1 ) 0 ÁÁÁ 0 0 6 7 0 0 7 f (l2 ) Á Á Á 6 0 6 7 6 . .47 Find f (A) ¼ Ak if 0:75 A ¼ 4 0:5 0:5 2 3 0:5 À0:75 3 À3 5 2 À2 1 X k¼0 ak (MLM À1 )k ¼ 1 X k¼0 ak MLk M À1 ¼ M 1 X k¼0 ak Lk M À1 (3:189) 2 1 P ak lk 1 3 7 7 7 7 0 7 7 7 7 . . . . . 4 0 0 0 0 Example 3. 7 7 7 0 7 7 7 1 5 P k ak ln 0 k¼0 (3:190) 0 ÁÁÁ 0 f (ln ) 3 (3:191) 2 Á Á Á f (lnÀ1 ) ÁÁÁ 0 ÁÁÁ ÁÁÁ . .. .

Mathematical Foundations

163

SOLUTION

The characteristic polynomial of A is l À 0:75 P(l) ¼ jlI À Aj ¼ À0:5 À0:5 À0:5 lÀ3 À2 0:75 3 ¼ l3 À 1:75l2 þ 0:875l À 0:125 lþ 2

¼ (l À 1)(l À 0:5)(l À 0:25) The eigenvalues are l1 ¼ 1, l2 ¼ 0:5, and eigenvectors are 2 3 0 x 1 ¼ 4 3 5, 2 Therefore, 2 0 1 1 32 (1)k 0 0 32 0 1 1 3À1 2 3 1 x2 ¼ 4 1 5, 1 l3 ¼ 0:25. The corresponding 2 3 1 x3 ¼ 4 2 5 2

and

7 76 76 6 f (A) ¼ Mf (L)MÀ1 ¼ 6 3 1 2 76 0 (0:5)k 0 76 3 1 2 7 5 54 54 4 2 1 2 2 1 2 0 0 (0:25)k 3 32 2 0 1 À1 0 (0:5)k (0:25)k 7 76 6 2 À3 7 ¼ 6 3 (0:5)k 2(0:55)k 76 2 5 54 4 k k À1 À2 3 2 (0:5) 2(0:25) 3 2 k k 2(0:5) À (0:25) 2(0:5)k À 2(0:25)k À3(0:5)k þ 3(0:25)k 7 6 ¼ 6 2(0:5)k À 2(0:25)k 3 þ 2(0:5)k À 4(0:25)k À3 À 3(0:5)k þ 6(0:25)k 7 5 4 2(0:5)k À 2(0:25)k Hence, 2 6 Ak ¼ 4 2(0:5)k À 2(0:25)k 2(0:5)k À 2(0:25)k 2(0:5)k À (0:25)k 2(0:5)k À 2(0:25)k 3 þ 2(0:5)k À 4(0:25)k 2 þ 2(0:5)k À 4(0:25)k 7 À3 À 3(0:5)k þ 6(0:25)k 5 À2 À 3(0:5)k þ 6(0:25)k À3(0:5)k þ 3(0:25)k 3 2 þ 2(0:5)k À 4(0:25)k À2 À 3(0:5)k þ 6(0:25)k

**3.11.5 MATRIX EXPONENTIAL FUNCTION eAt
**

The matrix exponential function eAt is a very important function with applications in analysis and design of continuous time-control systems. It is deﬁned as an inﬁnite series: eAt ¼ I þ At þ A2 t 2 A3 t 3 An t n þ þ ÁÁÁ þ þ ÁÁÁ 2! 3! n! (3:192)

164

Control of Color Imaging Systems: Analysis and Design

This series is a convergent series for all values of t and can be computed using different techniques such as Cayley–Hamilton and Laplace transform. The derivative of eAt with respect to t is deAt 2A2 t 2 A3 t 2 An t nÀ1 ¼Aþ þ þ ÁÁÁ þ þ ÁÁÁ dt 1! 2! (n À 1)! A2 t 2 A3 t 3 þ þ ÁÁÁ ¼ A I þ At þ 2! 3! ¼ AeAt ¼ eAt A The integral of the exponential function eAt is

t ð

(3:193)

eAt dt ¼ AÀ1 (eAt À I)

(3:194)

0

The above integral is valid if and only if matrix A is nonsingular. If A is singular, there is no simple closed-form solution.

Example 3.48

Find Ðt

0

eAt dt if 0 À2 2 1 1 À3 ! À2 À1 !

(a) A ¼ (b) A ¼

SOLUTION

(a) First we ﬁnd eAt using Cayley–Hamilton technique: l jlI À Aj ¼ 2 À1 ¼ (l þ 1)(l þ 2) ¼ 0 ! l1 ¼ À1, l2 ¼ À2 lþ 3

The corresponding eigenvectors are Ax1 ¼ l1 x1 , Ax2 ¼ l2 x2 , therefore x1 ¼ ½ 1 À1 T therefore x2 ¼ ½ 1 À2 T

Mathematical Foundations Therefore, eAt ¼ ML(t)MÀ1 ¼ ¼ 1 À1 1 À2 ! eÀt 0 0 ! 1 À1 1 À2 !

165

2eÀt À eÀ2t À2eÀt þ 2eÀ2t

eÀt À eÀ2t ÀeÀt þ 2eÀ2t

eÀ2t !

**Since matrix A is nonsingular, we have
**

t ð

eAt dt ¼ AÀ1 (eAt À I) ¼

0

! ! 1 0 2eÀt À eÀ2t eÀt À eÀ2t À 0 1 À2eÀt þ 2eÀ2t ÀeÀt þ 2eÀ2t ! ! Àt À2t Àt À2t À1:5 À0:5 2e À e À 1 e Àe ¼ Àt À2t Àt 1 0 À2e þ 2e Àe þ 2eÀ2t À 1 ! Àt À2t Àt À2e þ 0:5e þ 1:5 Àe À 0:5eÀ2t ¼ 2eÀt À eÀ2t À 1 eÀt À eÀ2t 0 1 À2 À3

!À1

(b) In this case, matrix A is singular, and eAt ¼ À1 þ 2et À1 þ et 2 À 2et 2 À et !

**Therefore, we use direct integration:
**

t ð

**3 Ðt ! (À1 þ 2et )dt (2 À 2et )dt 7 6 Àt À 2 þ 2et 2t þ 2 À 2et 7 6 0 eAt dt ¼ 6 0 t 7¼ Ðt Àt À 1 þ et 2t þ 1 À et 5 4Ð (À1 þ et )dt (2 À et )dt 0 2 Ðt
**

0 0

**3.11.6 COMPUTING eAt USING LAPLACE TRANSFORM
**

Laplace transform can be used to compute eAt . In this method, eAt is computed by solving the DE describing eAt using the Laplace transform. Let w(t) ¼ eAt , then dw(t) ¼ AeAt ¼ Aw(t) dt (3:195)

with w(0) ¼ eA0 ¼ I Taking the Laplace transform from both sides of Equation 3.195, we have sF(s) À w(0) ¼ AF(s) Since w(0) ¼ I, then (sI À A)F(s) ¼ I (3:197) (3:196)

166

Control of Color Imaging Systems: Analysis and Design

**and F(s) ¼ (sI À A)À1 Therefore, w(t) ¼ LÀ1 [(sI À A)À1 ] Example 3.49
**

Find eAt if A¼ À3 0 À1 À4 !

(3:198)

(3:199)

SOLUTION

Using Laplace transform, we have F(s) ¼ (sI À A)À1 ¼ or 3 sþ4 À1 6 (s þ 3)(s þ 4) (s þ 3)(s þ 4) 7 7 F(s) ¼ 6 5 4 sþ3 0 (s þ 3)(s þ 4) 2 Partial fraction of the entries of matrix F(s) yields 3 1 1 1 À þ 6s þ 3 s þ 3 s þ 47 7 F(s) ¼ 6 5 4 1 0 sþ4 2 Hence, eAt ¼ LÀ1 (sI À A)À1 ¼ eÀ3t 0 ÀeÀ3t þ eÀ4t eÀ4t ! sþ3 0 1 sþ4 !À1 ¼ 1 sþ4 0 s2 þ 7s þ 12 À1 sþ3 !

**3.11.7 MATRIX EXPONENTIAL FUNCTION Ak
**

Matrix exponential Ak has applications in design of discrete time-control systems. It is the state-transition matrix of LTI discrete systems in state-space form. It can be

Mathematical Foundations

167

computed using either Cayley–Hamilton technique or z-transform. To use z-transform, let w(k) ¼ Ak Then it is obvious that w(k þ 1) ¼ Akþ1 ¼ Aw(k) w(0) ¼ I Taking the z-transform from both sides of Equation 3.201, we have zF(z) À zw(0) ¼ AF(z) Since w(0) ¼ I, then we have (zI À A)F(z) ¼ zI Solving Equation 3.203 for F(z), F(z) ¼ z(zI À A)À1 Therefore, w(k) ¼ Z À1 [z(zI À A)À1 ] Example 3.50

Find Ak , if 0:1 A ¼ 4 À0:15 À0:2 2 3 0:4 À0:1 0:6 À0:05 5 0:4 0:2

(3:200)

(3:201)

(3:202)

(3:203)

(3:204)

(3:205)

SOLUTION

Using z-transform, we have z À 0:1 À0:4 ¼ z4 0:15 z À 0:6 0:2 À0:4 2 3À1 0:1 0:05 5 z À 0:2

F(z) ¼ z(zI À A)

À1

168

Control of Color Imaging Systems: Analysis and Design

**After taking the inverse, we have
**

F(z) ¼ z (z À 0:2)(z À 0:3)(z À 0:4) 2 (z À 0:6)(z À 0:2) þ 0:02 0:4(z À 0:2) À 0:04 6 6 À0:15(z À 0:2) þ 0:01 (z À 0:1)(z À 0:2) À 0:02 Â4 À0:2(z À 0:6) À 0:02 0:4(z À 0:1) À 0:08

7 À0:05(z À 0:1) À 0:015 7 5 (z À 0:1)(z À 0:6) þ 0:06

À0:1(z À 0:4) À 0:02

3

**Performing partial-fraction expansion, we have
**

3 z z z 2z 2z 2z 3z z þ À À þ À þ À 6 z À 0:2 z À 0:3 z À 0:4 z À 0:2 z À 0:4 z À 0:2 z À 0:3 z À 0:4 7 7 6 7 6 7 6 0:5z 0:5z z z 2z 0:5z 0:5z 7 F(z) ¼ 6 þ À À þ À 7 6 z À 0:2 z À 0:3 z À 0:4 z À 0:2 z À 0:4 z À 0:2 z À 0:3 7 6 7 6 4 3z 4z z 2z 2z 5z 12z 6z 5 À þ À þ À þ À z À 0:2 z À 0:3 z À 0:4 z À 0:2 z À 0:4 z À 0:2 z À 0:3 z À 0:4 2

Therefore,

w(k) ¼ Ak 2 3

6 ¼ 4 0:5(0:2)k þ 0:5(0:3)k À (0:4)k 3(0:2) À 4(0:3) þ (0:4)

k k k

(0:2)k þ (0:3)k À (0:4)k

À2(0:2)k þ 2(0:4)k À(0:2)k þ 2(0:4)k À2(0:2)k þ 2(0:4)k

7 5 0:5(0:2)k À 0:5(0:3)k k k k À5(0:2) þ 12(0:3) À 6(0:4)

À2(0:2)k þ 3(0:3)k À (0:4)k

**3.12 FUNDAMENTALS OF MATRIX CALCULUS
**

In a variety of multivariate optimization problems, it is necessary to differentiate a scalar function with respect to a vector. In this section, we look at derivatives of functions with respect to vectors.

3.12.1 DERIVATIVES

OF A

SCALAR FUNCTION

WITH

RESPECT TO

A

VECTOR

Let f (x) be a scalar function of vector x 2 Rn , then by deﬁnition, the derivative or gradient of f (x) with respect to vector x is an n Â 1 vector, which is deﬁned as 3 qf (x) 6 qx1 7 6 7 6 7 6 qf (x) 7 6 qf (x) 6 qx2 7 7 ¼6 7 6 . 7 qx 6 . 7 6 . 7 6 7 4 qf (x) 5 qxn 2

(3:206)

Mathematical Foundations

169

The second derivative of f (x) with respect to x 2 Rn is an n Â n matrix, which is deﬁned by q2 f (x) 6 qx2 6 1 6 6 2 6 q f (x) 6 2 q f (x) 6 qx2 qx1 ¼6 6 qx2 6 . 6 . 6 . 6 6 2 4 q f (x) qxn qx1 2 qf (x) qx1 qx2 q2 f (x) qx2 2 . . . q2 f (x) qxn qx2 3 q2 f (x) ÁÁÁ qx1 qxn 7 7 7 7 2 q f (x) 7 7 ÁÁÁ qx2 qxn 7 7 7 . 7 . . 7 . . 7 . 7 7 2 q f (x) 5 ÁÁÁ qx2 n

(3:207)

Example 3.51

Let x 2 R3 and f (x) ¼ x3 þ 3x2 þ 2x2 À 7x2 x3 þ 4. Find 1 1

qf (x) qx .

SOLUTION

3 qf (x) 6 qx1 7 2 2 3 7 6 3x1 þ 6x1 7 6 7 6 qf (x) 7 6 qf (x) 6 7 ¼ 6 2 À 7x3 7 ¼6 7 4 5 qx 6 qx2 7 7 6 À7x2 4 qf (x) 5 2 qx3 The second derivative is 2 q f (x) 6 ¼4 qx2

2

6x1 þ 6 0 0

0 0 À7

7 À7 5 0

0

3

Example 3.52

Let x 2 Rn and f (x) ¼ cT x, where c is an n Â 1 column vector, then f (x) ¼ cT x ¼ c1 x1 þ c2 x2 þ Á Á Á þ cn xn

170 and

Control of Color Imaging Systems: Analysis and Design

3 qf (x) 6 qx1 7 2 3 7 6 c1 7 6 6 qf (x) 7 6 7 qf (x) 6 qx 7 6 c2 7 7 6 7 6 ¼6 2 7¼6 . 7 6 . 7 6 . 7 qx 6 . 7 4 . 5 6 . 7 7 6 cn 4 qf (x) 5 2 qxn Hence, qcT x ¼c qx Similarly, qxT c ¼c qx (3:209) (3:208)

3.12.2 DERIVATIVES

OF

QUADRATIC FUNCTIONS

Let x 2 Rn and f (x) ¼ xT Ax, where A is an n Â n symmetric matrix. The multivariate quadratic function f (x) can be written as f (x) ¼ xT Ax ¼ Consider the ﬁrst component of vector

n n XX i¼1 j¼1 qf (x) qx ,

aij xi xj

(3:210)

that is,

n n n n qf (x) q XX q XX q ¼ aij xi xj ¼ aij xi xj þ a11 x1 x1 qx1 qx1 i¼1 j¼1 qx1 i¼2 j¼2 qx1

þ ¼

n n q X q X a1j x1 xj þ ai1 xi x1 qx1 j¼2 qx1 i¼2

n n n n X X q XX q q q aij xi xj þ a11 x2 þ x1 a1j xj þ x1 ai1 xi (3:211) 1 qx1 i¼2 j¼2 qx1 qx1 j¼2 qx1 i¼2

The ﬁrst term in the above expansion is independent of x1 , therefore it does not contribute to the derivative. Hence,

n n n n n X X X X X qf (x) ¼ 0 þ 2a11 x1 þ a1j xj þ ai1 xi ¼ a1j xj þ ai1 xi ¼ 2 a1j xj qx1 j¼2 i¼2 j¼1 i¼1 j¼1

(3:212)

Mathematical Foundations

171

This is equal to the ﬁrst element of the product 2Ax. Similarly, it can be shown that qf (x) qxi is equal to the ith element of the product. Therefore, qxT Ax ¼ 2Ax qx Similarly, it can be shown that the second derivative of f (x) is given by q2 xT Ax ¼ 2A qx2 Example 3.53

f Let x 2 R3 and f (x) ¼ 13x2 þ 8x2 À 4x2 À 3x1 x2 þ 5x1 x3 þ 6x2 x3 . Find qf (x) and qqx(x). 2 1 2 3 qx

2

(3:213)

(3:214)

SOLUTION

f (x) ¼ 13x2 þ 8x2 À 4x2 À 3x1 x2 þ 5x1 x3 þ 6x2 x3 1 2 3 32 3 2 x1 13 À1:5 2:5 76 7 6 ¼ ½ x1 x2 x3 4 À1:5 8 3 54 x2 5 ¼ xT Ax 2:5 3 À4 x3 The ﬁrst derivative of f (x) is 2 13 qf (x) ¼ 2Ax ¼ 24 À1:5 qx 2:5 The second derivative of f (x) is 2 26 q2 f (x) q2Ax ¼ ¼ 2A ¼ 4 À3 qx2 qx 5 3 À3 5 16 6 5 6 À8 32 3 2 3 x1 26x1 À 3x2 þ 5x3 À1:5 2:5 8 3 54 x2 5 ¼ 4 À3x1 þ 16x2 þ 6x3 5 3 À4 x3 5x1 þ 6x2 À 8x3

Example 3.54

Let x 2 Rn , A be an n Â n symmetric matrix, b be an n Â 1 column vector, and c a 2 f scalar. Deﬁne f (x) ¼ 1 xT Ax þ bT x þ c. Find qf (x) and qqx(x). 2 qx 2

SOLUTION

The ﬁrst derivative of f (x) is qf (x) ¼ Ax þ b qx

172

Control of Color Imaging Systems: Analysis and Design

And the second derivative is q2 f (x) q(Ax þ b) ¼ ¼A qx2 qx

3.12.3 DERIVATIVE

m

OF A

VECTOR FUNCTION

2

WITH

n

RESPECT

TO A

VECTOR

Let f (x) 2 R be a vector function of vector x 2 R , then by deﬁnition qf1 (x) qf2 (x) 6 qx1 qx1 6 6 6 qf1 (x) qf2 (x) qf (x) 6 qx2 6 qx2 ¼6 6 qx . 6 . . 6 . . 6 . 4 qf1 (x) qf2 (x) qxn qxn Example 3.55

Let x 2 R3 and f (x) ¼ 2

3 qfm (x) qx1 7 7 7 qfm (x) 7 7 ÁÁÁ qx2 7 7 7 . 7 . . 7 . . 7 . qfm (x) 5 ÁÁÁ qxn ÁÁÁ

(3:215)

! x2 x2 þ 3x2 þ 4x3 À 5 1 , then x1 x2 þ x2 À 2x1 x2 x3 3 qf2 (x) 3 qx1 7 2 2x x 7 x2 À 2x2 x3 1 2 7 7 6 2 7 qf2 (x) 7 6 7 7 ¼ 4 x1 þ 3 x1 À 2x1 x3 þ 1 5 qx2 7 7 4 À2x1 x2 qf2 (x) 5 qx3

qf1 (x) 6 qx1 6 6 qf (x) 6 qf1 (x) ¼6 6 qx qx 2 6 6 4 qf1 (x) qx3

PROBLEMS

3.1 (a) Let x(n) ¼ d(n) þ 2d(n À 1) þ 3d(n À 2) h(n) ¼ d(n) þ d(n À 1) þ 2d(n À 2) Find x(n) * h(n). (b) Consider the LTI system with impulse response h(n) ¼ À1Án

2

and

u(n)

Mathematical Foundations

173

**Using z-transform, ﬁnd the response y(n) of this system to the input x(n) ¼ À1Án
**

4

u(n)

3.2 Determine the ROCs and z-transforms of each of the following sequences: (a) x(n) ¼ (0:3)n u(n) þ 2n u(Àn À 1) (b) y(n) ¼ (0:2)n u(n À 2) (c) z(n) ¼ u(n) þ d(n) þ 3n u(Àn) 3.3 Evaluate the inverse z-transform of the following function: (a) X(z) ¼ z2 À z þ 1 6 , z2 À 5 z þ 1 6 6 ROC:

1 3

<z<1 2

(b) X(z) ¼ log (1 À 2zÀ1 ),

ROC: jzj > 2

3.4 Determine the z-transform and its ROC for each of the following 2-D sequences: (a) x(n1 , n2 ) ¼ ( 1 )n1 u(n1 )u(n2 ) 2 (b) y(n1 , n2 ) ¼ ( 1 )n1 þn2 u(n1 )u(n2 ) 3 3.5 Find the eigenvalues and eigenvectors of matrix A given below: 1 A¼ 3 2 À1 !

3.6 Find the eigenvalues and eigenvectors of the following matrices: 2 A¼ 0 ! À1 , 3 a B¼ b ! Àb , a 0 C¼4 0 À3 2 1 3 0 3 3 05 0

and

3.7 Let A be a 4 Â 4 matrix with eigenvalues À1, À2, À3, and À4. Find (if possible) the following quantities: (a) det (AT ) (b) Trace (AÀ1 ) (c) det (A À 8I) 3.8 Compute the eigenvalues and eigenvectors of the matrix 0 A ¼ 42 2 2 3 2 2 0 25 2 0

Can matrix A be diagonalized? If yes, ﬁnd the transformation to diagonalize matrix A.

174

Control of Color Imaging Systems: Analysis and Design

3.9 Compute the eigenvalues and eigenvectors of the matrix 1 A ¼ 40 0 2 3 1 1 2 15 0 3

Find, if possible, the matrix M such that M À1 AM is a diagonal matrix. 3.10 Show that the eigenvalues of A þ aI are related to the eigenvalues of A by l(A þ aI) ¼ l(A) þ a 3.11 Find the SVD of the following matrices: 1 A¼ À2 ! À2 , 4 1 À1 B ¼ 4 3 À2 1 À3 2 3 0 0 5, and 5 1 C ¼ 42 3 2 3 À1 À3 5 4

3.12 Consider the symmetric square matrix b A¼ 1 1 a ! 4 0 0 À2 ! b 1 1 a !À1

(a) Find eigenvalues of A þ 3I. (b) For what value of a, the matrix A þ aI is singular. 3.13 Determine whether the following matrices are (a) positive deﬁnite, (b) positive semi-deﬁnite, (c) negative deﬁnite, and (d) negative semi-deﬁnite: 2 A ¼ 41 1 2 1 4 0 3 1 0 5, 1 À1 B ¼ 4 À1 À1 2 À1 À4 À1 3 À1 À1 5, À2 7 À6 C¼ 6 2 !

and

3.14 Consider the following square matrix: A¼ 2 1 À1 2 !

(a) Compute Ak and eAt using the Cayley–Hamilton theorem. (b) Find the square roots of the matrix, that is, ﬁnd all matrices B such that B2 ¼ A using the Cayley–Hamilton theorem method.

Mathematical Foundations

175

3.15 Consider the following square matrix: 2 A ¼ 40 0 2 3 2 0 0 25 0 2

(a) Compute A7 and A73 using the Cayley–Hamilton theorem. (b) Compute eAt using the Cayley–Hamilton theorem. 3.16 Assume that the following matrix is nonsingular: a C ¼ 40 0 2 1 b 0 3 0 05 c

Compute the natural logarithm of the matrix C; that is, identify the matrix B ¼ ln (C) that satisﬁes the equation C ¼ eB . Is the assumption of non-singularity of C really needed for the solution of this problem? 3.17 Given the 3 Â 3 matrix À3 A¼4 1 À2 2 3 2 4 0 À1 5 1 5

Express the matrix polynomial f (A) ¼ A4 þ 2A3 þ A2 À A þ 3I as a linear combination of the matrices A2 , A, and I only. 3.18 Let the 2 Â 2 matrix A be deﬁned as A¼ Compute Ak and eAt . 3.19 Let the 3 Â 3 matrix A be deﬁned as 0 A ¼ 4 À3 0 Compute Ak and eAt . 2 3 À3 0 0 À3 5 À3 0 0 7 À1 0 !

DeRusso. 5. L. 3. Willsky. 1996. Signals and Systems. A. H. M. New York. Close. 4. 1994. The Laplace Transform: Theory and Applications. Digital Image Processing: Principles and Applications. Wiley-Interscience. 2nd edn. Roy. Baxes. and S. S. J. V. 2 3 À1 À1 À1 À1 5 0 À1 REFERENCES 1. Introduction to Higher Algebra. P. State Variables for Engineers. J. 1999. M. G. 1997. New York. . John Wiley & Sons. Dover Publications. Schiff.. 2nd edn. and A. NJ. Desrochers.. Oppenheim. Prentice Hall. Springer-Verlag New York Inc. Upper Saddle River. New York.20 Let the 3 Â 3 matrix A be deﬁned as À1 A¼4 0 0 Compute Ak and eAt . A. 2. A.. M. Bocher. Nawab. R. 2004. C.176 Control of Color Imaging Systems: Analysis and Design 3.

time-invariant to time-varying systems. In general terms. then these N variables are considered as components of an N-dimensional vector x called the state vector. 4.2 CONCEPT OF STATES The state-space approach is a uniﬁed approach for representation of both continuousand discrete-time dynamical systems.3 STATE-SPACE REPRESENTATION OF CONTINUOUS-TIME SYSTEMS 4. In this chapter. and controllability and observability of linear timeinvariant (LTI) systems. The state of a system at time t is a point in the N-dimensional state space [2.1 INTRODUCTION Modem control system theory is based on the state-space formulation of an underlying system. 177 . It covers a broad range of systems from nonlinear to linear. The N-dimensional space whose coordinates are the states of the system is called state space.4 State-Variable Representation 4. 4. If N variables are needed.3]. It can also be used in modeling stochastic dynamic systems.1 DEFINITION OF STATE The state of a continuous-time dynamic system is the minimum number of variables called state variables such that the knowledge of these variables at time t ¼ t0 together with the input for t ! t0 uniquely determines the behavior of the system for t ! t0 . time-varying and time-invariant systems as well as single-input single-output (SISO) and multiple-input multipleoutput (MIMO) systems [1]. solution of state equations. the states of a dynamic system is the minimum number of variables called state variables such that knowledge of these states at any given time together with the input at that time and future time uniquely determines the behavior of the system past that time [2]. we present system modeling in state space.3. It provides a uniﬁed approach for system modeling and design that is applicable to both linear and nonlinear systems.

. . . . u2 (t).3. u(t). . xN (t).178 Control of Color Imaging Systems: Analysis and Design OF 4. x2 (t). x2 (t). fN (x1 (t). uM (t). . .2 STATE EQUATIONS CONTINUOUS-TIME SYSTEMS The state-space equations describing a MIMO continuous-time dynamic system have a general form given by dx(t) ¼ f (x(t). 4 5 . u(t). and time-varying systems. uM (t). t) 7 6 7 h(x(t). then the state equations can be written as dx(t) ¼ A(t)x(t) þ B(t)u(t) dt y(t) ¼ C(t)x(t) þ D(t)u(t) (4:5) (4:6) where A(t). xN (t). .3 and 4. . . xN (t). . . u1 (t). u2 (t). . . x2 (t). u1 (t). t) 6 h2 (x1 (t). x2 (t). . . u2 (t). uM (t). hP (x1 (t). . . . 4 5 . If the system is LTI. . u(t). B(t).4 are applicable to both linear and nonlinear. . . u1 (t). . If the system is linear. . . xN (t). then A. . uM (t). u1 (t). t) and 3 h1 (x1 (t). u2 (t). . . C. xN (t). . uM (t). . t) ¼ 6 (4:3) 7 . and D are constant matrices and the state equation becomes dx(t) ¼ Ax(t) þ Bu(t) dt The output equation is y(t) ¼ Cx(t) þ Du(t) (4:8) (4:7) . . . t) 2 (4:1) (4:2) (4:4) The state equations given by Equations 4. B. u2 (t). C(t). u2 (t). . and D(t) are time-varying matrices of appropriate dimensions. . . u1 (t). x2 (t). x2 (t). u(t). . . t) dt y(t) ¼ h(x(t). . t) 7 6 7 f (x(t). . . t) ¼ 6 7 . . . t) where x(t) ¼ ½ x1 (t) x2 (t) Á Á Á xN (t) T is the N Â 1 state vector u(t) is the M Â 1 input vector y(t) is the P Â 1 output vector The vector functions f (:) and h(:) are in general nonlinear functions of the state and input vectors and are given by the following equations: 2 3 f1 (x1 (t). . . . . . . t) 6 f2 (x1 (t). u1 (t). time-invariant. uM (t). xN (t). . .

The ﬁrst state is deﬁned as the voltage across the capacitor and the second state the current through the inductor. one needs to know the input signal. where A. and D are N Â N. the initial charge on the capacitor. 4. N Â M. and P Â M constant matrices. the system has two states. To determine the output of this system.3. respectively. C. we show how the state equations can be derived for electrical and mechanical systems. . In the next section.1.1 OF ELECTRICAL SYSTEMS Consider the series RLC circuit shown in Figure 4. P Â N.1 Block diagram of a continuous LTI system in state space.State-Variable Representation 179 D · x (t) u(t) B + x(t) + C + + y(t) A FIGURE 4. The block diagram of a linear continuous time-invariant control system in state-space form is shown in Figure 4. B.2 with the input u(t) and output y(t). the ﬁrst state is deﬁned as x1 (t) ¼ vC (t) and the second state is deﬁned as x2 (t) ¼ iL (t) (4:10) (4:9) R + iL(t) u(t) L + + vC (t) – C y(t) – – FIGURE 4.3 STATE-SPACE EQUATIONS Example 4.2 Series RLC circuit. Therefore. therefore. and the initial current through the inductor.

the capacitor and the inductor.9 yields _ x1 (t) ¼ dvC (t) iL (t) 1 ¼ ¼ x2 (t) dt C C (4:11) Similarly. the A. hence. the number of states needed to model an electric circuit in state-space form is equal to the total number of independent energy-storage elements in the circuit. C. B. C ¼ ½ 1 0 . two states are needed. Example 4. and D matrices of the system are A¼ 1 C (4:13) (4:14) À1 L ! 0 R . there are two elements that can store energy. we consider an electric circuit with three independent energy-storage elements. ÀL B¼ 0 1 L ! .2 Consider the circuit shown in Figure 4. The three states of the circuit are deﬁned as follows: x1 (t) ¼ vC2 (t) x2 (t) ¼ vC1 (t) x3 (t) ¼ iL (t) (4:16) .10: _ x2 (t) ¼ diL (t) vL (t) 1 1 ¼ ¼ [u(t) À vR (t) À vC (t)] ¼ [u(t) À Rx2 (t) À x1 (t)] dt L L L u(t) R 1 ¼ À x2 (t) À x1 (t) L L L (4:12) The output equation is given by y(t) ¼ x1 (t) The above equations can be written in matrix form as # ! ! " 1 ! 0 _ x1 (t) 0 x1 (t) ¼ C1 þ 1 u(t) _ À L À R x2 (t) x2 (t) L L ! x1 (t) y(t) ¼ ½ 1 0 x2 (t) Therefore.2. In Example 4.3. In the circuit shown in Figure 4. we differentiae both sides of Equation 4. and D¼0 (4:15) As it can be seen from this simple example.2.180 Control of Color Imaging Systems: Analysis and Design Differentiating both sides of Equation 4.

4.18.22 can be written in matrix form as 2 _ x1 (t) _ x3 (t) 3 2 À1 3 1 3 À1 2 1 3 À1 3 (4:17) (4:18) (4:19) (4:20) (4:21) (4:22) 7 6 6_ 4 x2 (t) 5 ¼ 4 2 3 0 7 6 7 76 0 54 x2 (t) 5 þ 4 0 5u(t) 1 À1 2 x1 (t) x3 (t) 1 2 32 3 (4:23) 0 .20. and 4. we have _ x2 (t) ¼ 1 x1 (t) À 1 x2 (t) 3 3 Kirchhoff’s current law (KCL) at Node 1 yields _ _ _ _ x3 (t) ¼ C1 x2 (t) þ C2 x1 (t) ¼ x2 (t) þ x1 (t) Therefore. Kirchhoff’s voltage law (KVL) around the second loop is _ x x1 (t) ¼ vC2 (t) ¼ vC1 (t) þ vR2 (t) ¼ x2 (t) þ R2 C1 x2 (t) ¼ x2 (t) þ 3_ 2 (t) Therefore. _ _ x1 (t) ¼ x3 (t) À x2 (t) ¼ À 1 x1 (t) þ 1 x2 (t) þ x3 (t) 3 3 Finally. we have _ _ x3 (t) ¼ À 1 x3 (t) À 1 x2 (t) À 3 x2 (t) þ 1 u(t) ¼ À 1 x1 (t) À 1 x3 (t) þ 1 u(t) 2 2 2 2 2 2 2 Equations 4.State-Variable Representation R1 = 1 Ω + L =2 H iL(t) C1 = 1 F – R2 = 3 Ω Second loop 181 Node 1 + vC1(t) C2 = 1 F – – + vC2(t) = y(t) + u(t) – FIGURE 4. writing the KVL around the ﬁrst loop results in _ x R1 x3 (t) þ L_ 3 (t) þ x2 (t) þ R2 x2 (t) ¼ u(t) Hence.3 Electric circuit with three states.

Applying Newton’s law. Example 4. The new state and output equations are given by dz(t) ¼ PAPÀ1 z(t) þ PBu(t) dt y(t) ¼ CPz(t) þ Du(t) (4:27) 4. we have m d2 x(t) ¼ u(t) À kx(t) dt2 x=0 x k u(t) m (4:28) FIGURE 4. where m is the mass. .3 As the ﬁrst example. consider the ideal spring–mass system shown in Figure 4.4 Spring–mass system.3.4 STATE-SPACE EQUATIONS OF MECHANICAL SYSTEMS The state-space modeling of mechanical systems can be derived from the ﬁrst principles similar to the electrical circuits. and F is the overall forcing function with the assumption of linear spring. Any linear combinations of a set of states can be used as a new set of states for a given system.182 Control of Color Imaging Systems: Analysis and Design The output equation is given by y(t) ¼ x1 (t) ¼ ½ 1 0 3 x1 (t) 0 4 x2 (t) 5 x3 (t) 2 (4:24) Note that the states of a system are not unique. F ¼ ma. This means that if the N-dimensional vector x is a state of a dynamic system with state and output equations given by dx(t) ¼ Ax(t) þ Bu(t) dt y(t) ¼ Cx(t) þ Du(t) then the N-dimensional vector z is deﬁned as z ¼ Px (4:26) (4:25) where P is a nonsingular N Â N transformation.4. which is another state vector for the same system. The main equation in this case is Newton’s law of motion. a is the acceleration.

5 is the classical nonlinear control problem known as inverted pendulum. Then. The goal is to maintain the pendulum at the vertical position.5.State-Variable Representation 183 Deﬁne the two states of the system to be the mass displacement x(t) and its velocity _ x(t). the state equations become " _ x1 (t) _ x2 (t) # 3" # 2 3 0 1 0 x1 (t) ¼4 k þ 4 1 5u(t) b5 À À x2 (t) m m m 2 (4:33) Example 4. Example 4. It is assumed that the cart and the pendulum move in only one plane and the friction is negligible. The horizontal (Fx ) and vertical (Fy ) forces applied by the cart on the pendulum are given by Fx ¼ m d2 d _ _ (x þ l sin u) ¼ m (_ þ l u cos u) ¼ m(€ þ l € cos u À l u2 sin u) x x u dt2 dt (4:34) . consider the same ideal spring–mass system with friction that is assumed to be proportional to velocity.5 Consider a cart with an inverted pendulum as shown in Figure 4. we have x1 (t) ¼ x(t) _ x2 (t) ¼ x(t) Differentiating both sides of Equations 4.29 results in _ x1 (t) ¼ dx(t) ¼ x2 (t) dt 2 d x(t) k 1 k 1 _ ¼ À x(t) þ u(t) ¼ À x1 (t) þ u(t) x2 (t) ¼ dt2 m m m m (4:29) (4:30) The above equations can be written in matrix form as " _ x1 (t) _ x2 (t) # 3" # 2 3 0 1 0 x1 (t) 5 ¼4 k þ 4 1 5u(t) 0 À x2 (t) m m 2 (4:31) Example 4. Applying Newton’s law of motion to the system results in m d2 x(t) dx(t) ¼ u(t) À kx(t) À b dt2 dt (4:32) Using the same states.4 As another example.

184 Control of Color Imaging Systems: Analysis and Design θ m mg Fx u(t) M Fy x FIGURE 4.40 are used to solve for € and € The results are x u. Fy ¼ mg À m d2 d _ _ (l cos u) ¼ mg þ m (lu sin u) ¼ mg þ ml(€ sin u þ u2 cos u) u dt2 dt (4:35) Appling Newton’s law of motion to the cart results in M d2 x _ ¼ u(t) À Fx ¼ u(t) À m(€ þ l € cos u À lu2 sin u) x u dt2 (4:36) Equation 4.5 Inverted pendulum.36 can be written as _ (M þ m)€ þ ml€ cos u À ml u2 sin u ¼ u(t) x u Newton’s law applied to the motion of the pendulum yields u ml 2 € ¼ mgl sin u þ Fy l sin u À Fx l cos u Substituting Fx and Fy into Equation 4.38 results in _ u u ml 2 € ¼ mgl sin u þ m(g þ l € sin u þ l u2 cos u)l sin u _ À l cos u m(€ þ l u cos u À lu2 sin u) x € This equation can be simpliﬁed as _ x l(1 þ cos 2u)€ À l( sin 2u)u2 þ cos u€ ¼ 2g sin u u Equations 4.38 and 4. €¼ x Àmg sin 2u 1 þ cos 2u þ u(t) (2M þ m) cos u (2M þ m) cos u (4:41) (4:40) (4:38) (4:37) (4:39) .

then we have _ x1 ¼ x2 Àmg sin 2x3 1 þ cos 2x3 _ þ u(t) x2 ¼ (2M þ m) cos x3 (2M þ m) cos x3 _ x3 ¼ x4 2g(M þ m) sin x3 1 _ À x4 ¼ x2 sin x3 þ u(t) 4 (2M þ m)l cos x3 2M þ m (4:43) The above state equations are nonlinear. We deﬁne two useful canonical forms that are suitable . x3 ¼ u.44 can be written in matrix form as 2 6 7 6 6 _ 6 x2 (t) 7 6 0 0 7 6 6 6 x (t) 7 ¼ 6 0 0 4 _3 5 6 4 _ x4 (t) 0 0 _ x1 (t) 3 2 0 1 0 2mg À 2M þ m 0 2g(M þ m) 2M þ m 3 3 2 3 0 2 0 x1 (t) 7 7 6 2 7 0 76 x (t) 7 6 76 2 7 6 2M þ m 7 76 7u(t) 7þ6 7 1 76 x3 (t) 7 6 0 74 7 5 6 5 5 4 1 x4 (t) À 0 2M þ m (4:45) 4. x2 ¼ x.State-Variable Representation and 1 € ¼ ( sin u)u2 þ 2g(M þ m) sin u À _ u u(t) (2M þ m)l cos u 2M þ m 185 (4:42) _ _ Deﬁne the state variables as x1 ¼ x. and x4 ¼ u. we linearize them about u ¼ 0.4 STATE-SPACE REPRESENTATION OF GENERAL CONTINUOUS LTI SYSTEMS Consider a SISO system described by the Nth order constant coefﬁcients differential equation with input u(t) and output y(t): dN y(t) dNÀ1 y(t) dN u(t) dNÀ1 u(t) þ a1 þ Á Á Á þ aN y(t) ¼ b0 þ b1 þ Á Á Á þ bN u(t) dt N dt NÀ1 dt N dt NÀ1 (4:46) The state-space representation of the system deﬁned by differential equation given in Equation 4. The results of linearization are _ x1 ¼ x2 À2mg 2 _ x2 ¼ x3 þ u(t) 2M þ m 2M þ m _ x3 ¼ x4 2g(M þ m) 1 _ x4 ¼ x3 À u(t) (2M þ m)l 2M þ m (4:44) The linear equation given by Equation 4. To obtain linear equations.46 is not unique.

7 6 6 7 4 xNÀ1 (t) 5 xN (t) (4:48) x1 (t) 3 _ x1 (t) _ x2 (t) . .4. 6 7¼6 . 76 . 7 6 . 7 þ . . . 7 þ b0 u(t) . . 76 . . . 54 x (t) 5 4 .186 Control of Color Imaging Systems: Analysis and Design for state feedback and state estimator design. 76 .4. 6 7 4 . . . . 3 2 3 2 32 3 0 Á Á Á 0 ÀaN bN À a N b 0 x1 (t) 0 Á Á Á 0 ÀaNÀ1 76 x2 (t) 7 6 bNÀ1 À aNÀ1 b0 7 7 6 76 7 1 Á Á Á 0 aNÀ3 76 . 4 xNÀ1 (t) 5 0 0 0 _ ÀaN ÀaNÀ1 ÀaNÀ2 _ xN (t) with the output equation 2 6 7 6 x2 (t) 7 6 7 6 . 7 þ 6 bNÀ2 À aNÀ2 b0 7u(t) (4:49) 76 . u(t) (4:47) . .1 CONTROLLABLE CANONICAL FORM The controllable canonical form is given by 2 0 1 0 6 7 6 0 0 1 6 7 6 . 0 7 61 7 6 7 60 7¼6 7 6. 4. The ﬁrst one is controllable canonical form and the second one is observable canonical form. . 6 7 6 . 3 2 32 x (t) 3 2 3 0 0 1 0 76 x2 (t) 7 6 0 7 6 7 6 7 . . . . . NÀ1 xN (t) 0 0 Á Á Á 1 Àa1 b1 À a1 b 0 6 6 6 6 6 4 xNÀ1 _ _ xN (t) with the output equation 2 y(t) ¼ ½ 0 0 0 6 7 6 7 6 7 Á Á Á 1 6 7 þ b0 u(t) 6 7 4 xNÀ1 (t) 5 xN (t) x1 (t) x2 (t) . .2 OBSERVABLE CANONICAL FORM The observable canonical form is given by 2 _ x1 (t) _ x2 (t) . (t) 5 4 . . 76 . 7 y(t) ¼ ½ bN À aN b0 bNÀ1 À aNÀ1 b0 bNÀ2 À aNÀ2 b0 ÁÁÁ b1 À a1 b0 6 . 7 6 7 7 . 4. 7 . 7 6 . . 7 Á Á Á 1 54 xNÀ1 (t) 5 4 0 5 Á Á Á Àa1 1 xN (t) ÁÁÁ ÁÁÁ . 7 6 . . . . 5 . 3 (4:50) . .

the transfer matrix is P Â M.54 yields sX(s) ¼ AX(s) þ BU(s) Hence.State-Variable Representation 187 Example 4.6 Consider the following SISO system: d3 y(t) d2 y(t) dy(t) d2 u(t) du(t) þ5 þ2 þ7 þ 8y ¼ 9 þ 13u(t) 3 2 dt dt dt dt2 dt (4:51) Obtain a state-space representation of this system in (a) controllable and (b) observable canonical forms. the transfer function is 1 Â 1 and for a MIMO system with M inputs and P outputs. X(s) ¼ (sI À A)À1 BU(s) The output equation in Laplace transform domain is Y(s) ¼ CX(s) þ DU(s) ¼ [C(sI À A)À1 B þ D]U(s) (4:57) (4:56) (4:55) (4:54) . Let the state equations of a MIMO system be dx(t) ¼ Ax(t) þ Bu(t) dt Taking Laplace transform from both sides of Equation 4.4. For a SISO system. SOLUTION (a) Controllable canonical form: 3 2 _ 0 1 x1 (t) 6_ 7 6 4 x2 (t) 5 ¼ 4 0 0 2 _ x3 (t) 3 2 3 3 2 x1 (t) x1 (t) 0 76 7 6 7 7 6 1 54 x2 (t) 5 þ 4 0 5u(t) and y(t) ¼ ½ 13 7 9 4 x2 (t) 5 (4:52) 0 x3 (t) 1 x3 (t) 32 À8 À2 À5 (b) Observable canonical form: 2 7 6 6_ 4 x2 (t) 5 ¼ 4 1 _ 0 x3 (t) _ x1 (t) 3 2 0 0 À8 1 À5 32 76 7 6 7 0 À2 54 x2 (t) 5 þ 4 7 5u(t) and y(t) ¼ ½ 0 0 x3 (t) 9 x1 (t) 3 2 13 3 2 7 6 1 4 x2 (t) 5 x3 (t) (4:53) x1 (t) 3 4.3 TRANSFER FUNCTION (MATRIX) FROM STATE-SPACE EQUATIONS The transfer function (matrix) can be derived from state and output equations.

H(s) ¼ ¼ 1 ½4 2 þ 6s þ 5 s 1 ½4 s2 þ 6s þ 5 ! ! sþ6 1 2 À5 s 3 ! 2s þ 15 23s þ 10 5 ¼ 2 s þ 6s þ 5 3s À 10 5 À1 s À1 5 5 sþ6 !À1 2 3 ! (4:60) (4:61) 4.62) and then the solution is substituted into algebraic equation (Equation 4. The homogeneous state equation is given by _ x(t) ¼ Ax(t) (4:64) .7 Find the transfer function corresponding to the state equations given by ! ! ! ! _ 0 1 x1 (t) 2 x1 (t) u(t) ¼ þ _ À5 À6 x2 (t) 3 x2 (t) and y(t) ¼ ½ 4 5 x1 (t) x2 (t) ! (4:59) (4:58) SOLUTION The transfer function is H(s) ¼ C(sI À A) B þ D ¼ ½ 41 Hence.5 SOLUTION OF LTI CONTINUOUS-TIME STATE EQUATIONS Consider an LTI system described by the state-space equations: _ x(t) ¼ Ax(t) þ Bu(t) with the output equation. y(t) ¼ Cx(t) þ Du(t) (4:63) (4:62) To obtain the output of this system for a given input and an initial state.63) in order to ﬁnd the output y(t).5. the transfer matrix is H(s) ¼ C(sI À A)À1 B þ D Example 4. The solution to the state equation has two parts. homogeneous solution and particular solution [4]. we ﬁrst solve the state equation (Equation 4. 4.188 Control of Color Imaging Systems: Analysis and Design Therefore. We ﬁrst consider the homogeneous solution.1 SOLUTION OF HOMOGENEOUS STATE EQUATION The homogeneous solution is the solution of state equations to an arbitrary initial condition with zero input.

These properties are stated without proof: (a) w(0) ¼ I (b) w(Àt) ¼ eÀAt ¼ (eAt )À1 ¼ wÀ1 (t) (c) w(t1 þ t2 ) ¼ w(t1 )w(t2 ) ¼ w(t2 )w(t1 ) (d) w(t3 À t2 )w(t2 À t1 ) ¼ w(t3 À t1 ) The proof is left as an exercise (see Problem 4. Here. To verify that this is the solution to the homogeneous equation. we need to show that it satisﬁes the initial condition as well as the differential equation. .State-Variable Representation 189 Assuming that the initial state is x(0).5). the homogeneous part of the solution is x(t) ¼ eAt x(0). then x(t) ¼ eAt x(0) (4:65) where eAt is the matrix exponential function.5.64. which are very useful. we review two methods. we have _ x(t) ¼ deAt x(0) ¼ AeAt x(0) ¼ Ax(t) dt (4:67) (4:66) Therefore. The matrix exponential eAt is called state-transition matrix and is denoted by w(t). The initial condition is satisﬁed since x(0) ¼ eA0 x(0) ¼ Ix(0) ¼ x(0) Differentiating both sides of Equation 4. The state-transition matrix is given by w(t) ¼ eAt (4:69) (4:68) The state-transition matrix satisﬁes the following properties.2 COMPUTING STATE-TRANSITION MATRIX There are several methods for computing the state-transition matrix. (4:70) (4:71) (4:72) (4:73) 4. The statetransition matrix w(t) is an N Â N matrix and is the solution to the homogeneous equation w(t) ¼ Aw(t) _ with initial condition w(0) ¼ I.

w(t) is computed using the Laplace transform. ÁÁÁ e ln t 3 0 0 7 . 5 . 4 . we have sF(s) À w(0) ¼ AF(s) Since w(0) ¼ I.. Example 4.8 Find the state-transition matrix of the following dynamic system using the two methods described above: ! À1 1 _ x(t) x(t) ¼ À1 À1 (4:76) (4:77) (4:78) SOLUTION (a) Modal Matrix Technique: The eigenvalues of matrix A are lþ1 jlI À Aj ¼ 1 1 ¼ (l þ 1)2 þ 1 ¼ 0 ! l1 ¼ À1 þ j. l2 ¼ À1 À j lþ 1 . w(t) ¼ LÀ1 [(sI À A)À1 ] where LÀ1 stands for inverse the Laplace transform. . 7 . . .62. 0 ÁÁÁ ÁÁÁ . the state-transition matrix is given by w(t) ¼ ML(t)M À1 (4:74) where M ¼ ½ x1 x2 Á Á Á xN is the N Â N matrix of eigenvectors known as modal matrix L(t) is an N Â N diagonal matrix given by e l1 t 6 0 L(t) ¼ 6 . Under the assumption of independent eigenvectors.190 Control of Color Imaging Systems: Analysis and Design (a) Modal Matrix Approach: Let {li . (b) Laplace Transform Technique: In this method. (4:75) The second approach which is more general is based on Laplace Transform. Taking the Laplace transform from both sides of Equation 4. . then F(s) ¼ (sI À A)À1 Therefore. xi }N be the set of eigenvalues and i¼1 their corresponding eigenvectors of matrix A. 0 2 e l2 t 0 .

we have w(t) ¼ eÀt cos t ÀeÀt sin t eÀt sin t eÀt cos t ! 4.62) is the sum of the homogeneous and particular solutions and is given by t ð x(t) ¼ w(t)x(0) þ w(t À t)Bu(t)dt 0 (4:79) . Hence. " w(t) ¼ ML(t)MÀ1 ¼ " ¼ " ¼ " ¼ 1 j 1 Àj #" eÀ(1Àj)t 0 eÀ(1þj)t ÀjeÀ(1þj)t þ 0:5e 0 #" 1 1 #À1 therefore x1 ¼ ½ 1 j T therefore x2 ¼ ½ 1 Àj T 191 eÀ(1Àj)t jeÀ(1Àj)t 0:5e j Àj eÀ(1þj)t # #" 0:5 À0:5j 0:5 0:5j À0:5jeÀ(1Àj)t þ 0:5jeÀ(1þj)t 0:5eÀ(1Àj)t þ 0:5eÀ(1þj)t # À(1Àj)t À(1þj)t 0:5jeÀ(1Àj)t À 0:5jeÀ(1þj)t # eÀt cos t eÀt sin t ÀeÀt sin t eÀt cos t (b) Use of Laplace Transform: Using Laplace transform. we have F(s) ¼ (sI À A)À1 ¼ or s þ 1 À1 1 sþ1 2 !À1 ¼ 1 sþ1 1 (s þ 1)2 þ 1 À1 s þ 1 ! sþ1 6 (s þ 1)2 þ 1 6 F(s) ¼ 6 4 À1 (s þ 1)2 þ 1 3 1 (s þ 1)2 þ 1 7 7 7 5 sþ1 (s þ 1)2 þ 1 Taking inverse Laplace transform.5.3 COMPLETE SOLUTION OF STATE EQUATION The complete solution of the state equation (Equation 4.State-Variable Representation The corresponding eigenvectors are Ax1 ¼ l1 x1 . Ax2 ¼ l2 x2 .

The second term is the particular solution. which is also referred to as zero-state response.79: d _ x(t) ¼ w(t)x(0) þ _ dt t ð w(t À t)Bu(t)dt 0 t ð ¼ Aw(t)x(0) þ w(t À t)Bu(t) þ w(t À t)Bu(t)dt _ 0 t ð ¼ Aw(t)x(0) þ w(0)Bu(t) þ Aw(t À t)Bu(t)dt 0 t ð ¼ Aw(t)x(0) þ Bu(t) þ A w(t À t)Bu(t)dt 2 0 t ð 3 ¼ A4w(t)x(0) þ w(t À t)Bu(t)dt5 þ Bu(t) 0 ¼ Ax(t) þ Bu(t) (4:81) The ﬁrst term in the total solution. Example 4. differentiate both sides of Equation 4. we need to show that it satisﬁes the state equation and the initial condition. that is w(t)x(0). u(t) ¼ n 1 0 t>0 t<0 and 1 x(0) ¼ 0 ! .9 Find the output of the dynamic system described by the state equation: " _ x(t) ¼ À1 1 # x(t) þ " # 1 0 u(t) À1 À1 À1 x(t) y(t) ¼ ½ 1 With the input and initial conditions given by.192 Control of Color Imaging Systems: Analysis and Design To verify that this is the total solution. The initial condition is satisﬁed since 0 ð x(0) ¼ w(0)x(0) þ w(0 À t)Bu(t)dt ¼ I Â x(0) þ 0 ¼ x(0) 0 (4:80) To show that it satisﬁes the state equation. is the homogeneous solution or the zero-input response.

If N variables are needed. then these N variables are considered as components of an N-dimensional vector x called state vector. The state of a system at time n is a point in the state space [5].1 DEFINITION OF STATE The state of a discrete-time dynamic system is the minimum number of variables called state variables such that the knowledge of these variables at time k ¼ k0 together with input for k ! k0 uniquely determines the behavior of the system for k ! k0 . The N-dimensional space whose coordinates are the states of the system is called state space.6. ! ! eÀt cos t À0:5eÀt cos t þ 0:5eÀt sin t þ 0:5 þ x(t) ¼ xh (t) þ xp (t) ¼ ÀeÀt sin t 0:5eÀt cos t À 0:5eÀt sin t À 0:5 ! 0:5eÀt cos t þ 0:5eÀt sin t þ 0:5 ¼ 0:5eÀt cos t À 1:5eÀt sin t À 0:5 The output y(t) is y(t) ¼ ½ 1 À1 x(t) ¼ ½ 1 À1 0:5eÀt cos t þ 0:5eÀt sin t þ 0:5 0:5eÀt cos t À 1:5eÀt sin t À 0:5 ! ¼ 2eÀt sin t þ 1 4.6 STATE-SPACE REPRESENTATION OF DISCRETE-TIME SYSTEMS 4.State-Variable Representation 193 SOLUTION The homogeneous solution or the zero-state response is given by ! ! ! 1 eÀt cos t eÀt cos t eÀt sin t ¼ xh (t) ¼ w(t)x(0) ¼ 0 ÀeÀt sin t eÀt cos t ÀeÀt sin t The zero-state response or the particular solution is t t ! ! ð ð Àtþt e cos(t Àt) eÀtþt sin(t Àt) 1 dt xp (t) ¼ w(t Àt)Bu(t)dt ¼ ÀeÀtþt sin(t Àt) eÀtþt cos(t Àt) 0 0 0 3 2 Ðt Ðt t eÀt cos t et cos t dtþeÀt sin t et sint dt 7 ! 6 ð Àtþt e cos(t Àt) 0 0 7 6 dt6 ¼ 7 5 4 Àt Ðt t Ðt t ÀeÀtþt sin(t Àt) Àt Àe sin t e cos t dtþe cos t e sin t dt 0 0 0 " # t Àt t t Àt t e cos t[0:5e cos tþ0:5e sin t]j0 þe sin t[À0:5e cos tþ0:5et sin t]jt 0 ¼ ÀeÀt sin t[0:5et cos tþ0:5et sin t]jt þeÀt cos t[À0:5et cos tþ0:5et sin t]jt 0 0 ¼ eÀt cos t[0:5et cos t þ0:5et sin t À0:5]þeÀt sin t[À0:5et cos t þ0:5et sin t þ0:5] ÀeÀt sin t[0:5et cos t þ0:5et sin t À0:5]þeÀt cos t[À0:5et cos t þ0:5et sin t þ0:5] ! À0:5eÀt cos t þ0:5eÀt sin t þ0:5 ¼ 0:5eÀt cos t À0:5eÀt sin t À0:5 ! Therefore. .

. uM (k). uM (k). B. . . . x2 (k). k) 7 6 7 h(x(n). . . . fN (x1 (k). . . u(k). . . x2 (k). . . u2 (k). u2 (k). k) 6 h2 (x1 (k). k) 7 6 7 f (x(k). uM (k). u2 (k). . . and D are N Â N. . k) y(k) ¼ h(x(k). .194 Control of Color Imaging Systems: Analysis and Design 4. 4 5 . u2 (k). . . . k) ¼ 6 7 .6. . . . k) (4:84) 2 3 h1 (x1 (k). u(n). N Â M. . . B.6. . . n) ¼ 6 7 . respectively.2 STATE EQUATIONS The general expressions for state-space equations of a MIMO dynamic system are given by x(k þ 1) ¼ f (x(k). k) where x(k) ¼ ½ x1 (k) x2 (k) Á Á Á xN (k) T is the N Â 1 state vector u(k) is the M Â 1 input vector y(n) is the P Â 1 output vector The functions f and h are nonlinear and are deﬁned as 3 f1 (x1 (k). uM (k). then the state equations can be written as x(k þ 1) ¼ A(k)x(k) þ B(k)u(k) y(k) ¼ C(k)x(k) þ D(k)u(k) (4:86) (4:87) 2 (4:82) (4:83) If the system is LTI. . . x2 (k). x2 (k). . k) (4:85) If the system is linear. uM (k). xN (k). k) 6 f2 (x1 (k). u1 (k). . . u1 (k). . u(k). . . . u1 (k). xN (k). C. . C. xN (k). . xN (k). . u1 (k). . . . . P Â N. u1 (k). xN (k). u2 (k). . x2 (k). and P Â M constant matrices. x2 (k). . . and D are constant matrices and the state equations become x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) þ Du(k) (4:88) (4:89) where A. u(k). u1 (k). u2 (k). . . . 4 5 . hP (x1 (k). . . The block diagram of an LTI control system in state-space form is shown in Figure 4. then matrices A. uM (k). xN (k).

are discussed below. . 6 7¼6 . 7 þ 6 . . . . 4. 7 ÁÁÁ 1 54 xNÀ1 (k) 5 4 0 5 1 ÁÁÁ Àa1 xN (k) ÁÁÁ ÁÁÁ . 7 6 .7. . 7 . 76 . . 7 6 4 xNÀ1 (k) 5 xN (k) 2 x1 (k þ 1) x2 (k þ 1) .e. 7 Á Á Á b1 À a1 b0 6 . 6 7 6 .State-Variable Representation 195 D x(k+1) u(k) B + + Unit delay x(k) C + + y(k) A FIGURE 4.. .7 STATE-SPACE REPRESENTATION OF DISCRETE-TIME LTI SYSTEMS Consider the LTI discrete-time dynamic system described by the difference equation: N X i¼1 M X i¼0 y(k) ¼ À ai y(k À i) þ bi u(k À i) (4:90) The state-space realizations of the above system in two most important standard forms. 6 7 4 4 xNÀ1 (k þ 1) 5 0 0 0 ÀaN ÀaNÀ1 ÀaNÀ2 xN (k þ 1) with the output equation 3 x1 (k) 6 x2 (k) 7 7 6 6 . 7 6 . . 4. 3 2 32 x (k) 3 2 3 0 0 1 0 76 x2 (k) 7 6 0 7 6 7 . 7 þ b0 u(k) . i. y(k) ¼ ½ bN À aN b0 bNÀ1 À aNÀ1 b0 bNÀ2 À aNÀ2 b0 (4:92) . . 76 . the controllable and observable canonical forms.1 CONTROLLABLE CANONICAL FORM The controllable canonical form is given by 2 0 1 0 6 7 6 0 0 1 6 7 6 . 76 . 7u(k) (4:91) .6 Block diagram of a discrete LTI system in state space. .

. Example 4. 7 þ 6 bNÀ2 ÀaNÀ2 b0 7u(k) (4:93) 76 . . . SOLUTION (a) Controllable canonical form: The state equations are ! 0 x1 (k þ 1) ¼ À0:25 x2 (k þ 1) and the output equation is y(k) ¼ ½ À4:25 2:25 1 À0:75 ! ! ! x1 (k) 0 u(k) þ 1 x2 (k) ! x1 (k) þ u(k) x2 (k) (b) Observable canonical form: The state equations and the output equation are ! 0 x1 (k þ 1) ¼ 1 x2 (k þ 1) ! ! x1 (k) À4:25 u(k) þ 2:25 x2 (k) ! x (k) y(k) ¼ ½ 0 1 1 þ u(k) x2 (k) À0:25 À0:75 ! and . 6 7 6 6 4 xNÀ1 (k þ1) 5 4 . 7¼6 6 . .7. . . . 7 6 76 7 . NÀ1 xN (k) 2 y(k) ¼ ½ 0 0 0 3 (4:94) b1 Àa1 b0 x1 (k) 3 2 bN ÀaN b0 3 0 0 ÁÁÁ 1 Àa1 with the output equation given as 6 7 6 7 6 7 Á Á Á 1 6 7 þ b0 u(k) 6 7 4 xNÀ1 (k) 5 xN (k) x1 (k) x2 (k) . . 76 . 54 x (k) 5 4 .196 Control of Color Imaging Systems: Analysis and Design 4.10 Consider the following SISO discrete-time LTI system y(k þ 2) þ 0:75y(k þ 1) þ 0:25y(k) ¼ u(k þ 2) þ 3u(k þ 1) À 4u(k) Obtain a state-space representation of this system in (a) controllable and (b) observable canonical forms. . . . 5 . xN (k þ1) x1 (k þ1) 3 2 0 0 ÁÁÁ 0 ÀaN 32 0 ÁÁÁ 0 ÀaNÀ1 76 x2 (k) 7 6 bNÀ1 ÀaNÀ1 b0 7 7 6 76 7 7 6 76 7 1 ÁÁÁ 0 aNÀ3 76 . 7 6 7 . .2 OBSERVABLE CANONICAL FORM The observable canonical form is given by 2 6 x (k þ1) 7 6 1 7 6 6 2 7 6 6 7 60 6 . . . 7 6. 7 6 7 .

the transfer matrix is P Â M. the transfer function is 1 Â 1 and for a MIMO system with M inputs and P outputs. the transfer function matrix is H(z) ¼ C(zI À A)À1 B þ D (4:99) (4:98) (4:97) (4:96) (4:95) Example 4.3 TRANSFER FUNCTION (MATRIX) STATE-SPACE EQUATIONS The transfer function (matrix) can be derived from state and output equations. For a SISO system.11 Find the transfer function corresponding to the state equation given by x1 (k þ 1) x2 (k þ 1) ! ¼ 0 1 ! x1 (k) ! þ 2 3 ! u(k) À0:6 À0:5 x2 (k) ! x1 (k) y(k) ¼ ½ 1 À2 x2 (k) SOLUTION The transfer function is H(z) ¼ C(zI À A)À1 B þ D ¼ ½ 1 or 1 z þ 0:5 1 H(z) ¼ 2 ½ 1 À2 À0:6 z z þ 0:5z þ 0:6 ! 2 3 ! À2 z 0:6 À1 z þ 0:5 !À1 2 3 ! .95 yields zX(z) ¼ AX(z) þ BU(z) Hence. Let the state equations of a discrete MIMO system be x(k þ 1) ¼ Ax(k) þ Bu(k) Taking z-transform from both sides of Equation 4. X(z) ¼ (zI À A)À1 BU(z) The output equation in z-domain is Y(z) ¼ CX(z) þ DU(z) ¼ [C(zI À A)À1 B þ D]U(z) Therefore.State-Variable Representation 197 FROM 4.7.

in matrix form. We ﬁrst consider the homogeneous solution.1 SOLUTION OF HOMOGENEOUS STATE EQUATION The homogeneous solution is the solution of state equations to an arbitrary initial condition with zero input. . we ﬁrst solve the state equation (Equation 4. Control of Color Imaging Systems: Analysis and Design H(z) ¼ À4z þ 6:4 z2 þ 0:5z þ 0:6 4. The matrix exponential Ak is called state-transition matrix and is denoted by w(k). The state-transition matrix w(k) is an N Â N matrix and is the solution to the homogeneous equation . a homogeneous solution and a particular solution. the homogeneous part of the solution is x(k) ¼ Ak x(0). The solution to state equations has two parts. we have x(1) ¼ Ax(0) x(2) ¼ Ax(1) ¼ A Â Ax(0) ¼ A2 x(0) x(3) ¼ Ax(2) ¼ A Â A2 x(0) ¼ A3 x(0) .101) in order to ﬁnd the output y(k). the solution to the homogeneous equation is x(k) ¼ Ak x(0) (4:104) (4:103) (4:102) Therefore. 4. x(k) ¼ Ax(k À 1) ¼ A Â AkÀ1 x(0) ¼ Ak x(0) Therefore. .100) and then the solution is substituted into the algebraic equation (Equation 4. The homogeneous state equation is given by x(k þ 1) ¼ Ax(k) Assuming that the initial state is x(0).8 SOLUTION OF LTI DISCRETE-TIME STATE EQUATIONS Consider a discrete LTI system described by the state-space equation x(k þ 1) ¼ Ax(k) þ Bu(k) and the output equation y(k) ¼ Cx(k) þ Du(k) (4:101) (4:100) To obtain the output of this system for a given input and an initial state.198 Therefore.8.

(a) w(0) ¼ I (b) w(Àk) ¼ AÀk ¼ (Ak )À1 ¼ wÀ1 (k) (c) w(k1 þ k2 ) ¼ w(k1 )w(k2 ) ¼ w(k2 )w(k1 ) (d) w(k3 À k2 )w(k2 À k1 ) ¼ w(k3 À k1 ) The proof is left as an exercise (see Problem 4. .State-Variable Representation 199 w(k þ 1) ¼ Aw(k) w(0) ¼ I The state-transition matrix is given by w(k) ¼ Ak (4:105) (4:106) The state transition matrix satisﬁes the following properties. . we discuss these two methods. Then the state-transition matrix is computed using w(k) ¼ ML(k)M À1 (4:111) where M ¼ ½ V1 V2 Á Á Á VN is the N Â N matrix of eigenvectors known as modal matrix L(k) is an N Â N diagonal matrix given by lk 1 60 6 L(k) ¼ 6 .8. 0 0 . 0 2 0 lk 2 . . 0 ÁÁÁ ÁÁÁ . the discrete state-transition matrix can be computed using modal matrix or z-transform.105. 3 7 7 7 5 (4:112) Á Á Á lk N (b) z-Transform Approach: In this method. (4:107) (4:108) (4:109) (4:110) 4. we have zF(z) À zw(0) ¼ AF(z) (4:113) . . Here. 4 .2 COMPUTING STATE-TRANSITION MATRIX Similar to the continuous case. . (a) Modal Matrix Approach: Let {li . . These properties are stated without proof. w(k) is computed using the z-transform. Vi }N be the set of eigenvalues and i¼1 their corresponding eigenvectors of matrix A..6). Taking the z-transform on both sides of Equation 4.

the eigenvalues of A are l1 ¼ 0:5 l2 ¼ 0:25 The corresponding eigenvectors are AV1 ¼ l1 V1 . AV2 ¼ l2 V2 . F(k) ¼ Z À1 [z(zI À A)À1 ] (4:116) (4:115) (4:114) Example 4.12 Find the state-transition matrix of the following dynamic system using the two methods presented previously: x(k þ 1) ¼ 1 1:5 ! À0:25 x(k) À0:25 SOLUTION (a) Modal Matrix Approach: The characteristic polynomial of matrix A is l À 1 0:25 2 jlI À Aj ¼ À1:5 l þ 0:25 ¼ l À 0:75l þ 0:125 ¼ (l À 0:5)(l À 0:25) ¼ 0 Therefore. w(k) ¼ ML(k)MÀ1 ¼ " ¼ 1 2 1 3 !" (0:5)k 0 # ! 1 1 À1 0 (0:25)k 2 3 # ! 3 À1 (0:25)k 3(0:25)k À2 1 (0:5)k 2(0:5)k . then (zI À A)F(z) ¼ zI and F(z) ¼ z(zI À A)À1 Therefore. therefore V1 ¼ ½ 1 therefore V2 ¼ ½ 1 2 T 3 T Using the modal matrix approach.200 Control of Color Imaging Systems: Analysis and Design Since w(0) ¼ I.

. we have F(z) ¼ z(zI À A)À1 ¼ z z z þ 0:25 À0:25 1:5 zÀ1 z2 À 0:75z þ 0:125 3 2 z þ 0:25 0:25 À 6 (z À 0:5)(z À 0:25) (z À 0:5)(z À 0:25) 7 7 F(z) ¼ z6 5 4 1:5 zÀ1 (z À 0:5)(z À 0:25) (z À 0:5)(z À 0:25) z À 1 0:25 À1:5 z þ 0:25 ¼ !À1 ! Partial fraction expansion yields 3 3 2 3z À2z Àz z 3 À2 À1 1 þ þ þ þ 6 zÀ0:5 zÀ0:25 zÀ0:5 zÀ0:25 7 6 zÀ0:5 zÀ0:25 zÀ0:5 zÀ0:25 7 7 7¼6 F(z)¼z6 4 6 À6z À2z 3z 5 À6 À2 3 5 4 6z þ þ þ þ zÀ0:5 zÀ0:25 zÀ0:5 zÀ0:25 zÀ0:5 zÀ0:25 zÀ0:5 zÀ0:25 2 Hence.3 COMPLETE SOLUTION OF STATE EQUATIONS Assuming that the initial state is x(0). x(k) ¼ Ak x(0) þ kÀ1 X n¼0 AkÀ1Àn Bu(n) (4:118) . 3(0:5)k À 2(0:25)k w(k) ¼ 6(0:5)k À 6(0:25)k À(0:5)k þ (0:25)k À2(0:5)k þ 3(0:25)k ! 201 (b) z-Transform Approach: Using z-transform. . we have x(1) ¼ Ax(0) þ Bu(0) x(2) ¼ Ax(1) þ Bu(1) ¼ A2 x(0) þ ABu(0) þ Bu(1) x(3) ¼ Ax(2) þ Bu(2) ¼ A3 x(0) þ A2 Bu(0) þ ABu(1) þ Bu(2) . x(k) ¼ Ax(k À 1) þ Bu(k À 1) ¼ Ak x(0) þ AkÀ1 Bu(0) þ AkÀ2 Bu(1) þ Á Á Á þ Bu(k À 1) (4:117) Therefore.8.State-Variable Representation Therefore. ! w(k) ¼ ZÀ1 [F(z)] ¼ 3(0:5)k À 2(0:25)k 6(0:5)k À 6(0:25)k À(0:5)k þ (0:25)k À2(0:5)k þ 3(0:25)k 4.

Example 4. the ﬁrst term w(k)x(0) is referred to as the PkÀ1 zero-input response and the second term n¼0 w(k À 1 À n)Bu(n) is called zero-state response.13 & Find the output of the following dynamic system if u(k) ¼ ! 1 x(0) ¼ : À1 " x(k þ 1) ¼ 1 À0:25 # x(k) þ " À1 1 # u(k) 1 k ! 0 and 0k<0 1:5 À0:25 y(k) ¼ ½ 1 2 x(k) SOLUTION First we ﬁnd the zero-input response: 3(0:5)k À 2(0:25)k À(0:5)k þ (0:25)k w(k)x(0) ¼ 6(0:5)k À 6(0:25)k À2(0:5)k þ 3(0:25)k Next. In the following example. The zero-input response is the response of the system due to the initial conditions only and the zero-state response is the response due to the input with zero initial conditions.202 Control of Color Imaging Systems: Analysis and Design In terms of state-transition matrix. we compute the zero-state and zeroinput responses of a second-order system. the total solution is given by x(k) ¼ w(k)x(0) þ kÀ1 X n¼0 w(k À 1 À n)Bu(n) (4:119) The total solution consists of two terms. we ﬁnd the zero-state response: kÀ1 X n¼0 ! ! ! 1 4(0:5)k À 3(0:25)k ¼ À1 8(0:5)k À 9(0:25)k w(k À 1 À n)Bu(n) ¼ " kÀ1 X w11 (k À 1 À n) n¼0 w12 (k À 1 À n) w22 (k À 1 À n) #" 1 À1 # ¼ kÀ1 X w11 (k À 1 À n) À w12 (k À 1 À n) " w21 (k À 1 À n) # ¼ w21 (k À 1 À n) À w22 (k À 1 À n) # kÀ1 X 4(0:5)kÀ1Àn À 3(0:25)kÀ1Àn n¼0 " n¼0 8(0:5)kÀ1Àn À 9(0:25)kÀ1Àn .

1 DEFINITION OF CONTROLLABILITY A discrete-time LTI system is said to be controllable at time k0 if there exists an input u(k) for k ! k0 that can transfer the system from any initial state x(k0 ) to the origin in ﬁnite number of steps.State-Variable Representation or 2 kÀ1 X kÀ1 203 2 À 3(0:25) 4 7 6 4(0:5) 7 6 n¼0 n¼0 7 w(k À 1 À n)Bu(n) ¼ 6 7 6 kÀ1 kÀ1 4 n¼0 kÀ1 P n kÀ1 P n 5 8(0:5) 2 À 9(0:25) 4 n n kÀ1 P kÀ1 kÀ1 P 3 2 n¼0 kÀ1 1 À 2 k n¼0 k kÀ1 1 À 4 3 À 3(0:25) 6 4(0:5) 1À2 1À4 7 7 6 ¼6 7 4 k k5 1À2 1À4 8(0:5)kÀ1 À 3(0:25)kÀ1 1À2 1À4 " # " # kÀ1 kÀ1 À4(0:5) þ 4 þ (0:25) 4 À 8(0:5)k þ 4(0:25)k ¼ ¼ À8(0:5)kÀ1 þ 12 þ (0:25)kÀ1 12 À 16(0:5)k þ 4(0:25)k The total system response is the sum of zero-state and zero-input responses. Since the results obtained are identical for continuous and discrete-time systems. Therefore.9 CONTROLLABILITY OF LTI SYSTEMS In general. we only consider controllability of discrete-time LTI systems [6.9. we have x(k) ¼ w(k)x(0) þ kÀ1 X n¼0 " w(k À 1 À n)Bu(n) ¼ " ¼ 4(0:5)k À 3(0:25)k 8(0:5)k À 9(0:25)k # þ " 4 À 8(0:5)k þ 4(0:25)k # 4 À 4(0:5)k þ (0:25)k 12 À 8(0:5)k À 5(0:25)k 12 À 16(0:5)n þ 4(0:25)k # The output signal is y(k) ¼ ½ 1 2 x(k) ¼ ½ 1 2 ! 4 À 5(0:5)k þ 2(0:25)k ¼ 28 À 25(0:5)k À 2(0:25)k 12 À 10(0:5)k À 2(0:25)k 4.7]. we examine the controllability of LTI systems. 4. a system is controllable if there exists an input that can transfer the states of the system from an arbitrary initial state to a ﬁnal state in a ﬁnite interval of time. In this section. .

B 6 7 ¼ ÀAN x(k0 ) . . . Example 4. .2 CONTROLLABILITY CONDITION The following theorem gives the necessary and sufﬁcient conditions for controllability of a discrete-time LTI system. u(k0 þ N À 1). we must have AN x(k0 ) þ ANÀ1 Bu(k0 ) þ ANÀ2 Bu(k0 þ 1) þ Á Á Á þ Bu(N þ k0 À 1) ¼ 0 or 3 u(N þ k0 À 1) 6 u(N þ k0 À 2) 7 7 Ã6 6 7 . .9.1 A discrete-time LTI system described by x(k þ 1) ¼ Ax(k) þ Bu(k) (4:120) is completely state controllable if the controllability matrix Q is full rank. Let the initial state be x(k0 ) at time k0 and the input for time k ! k0 be u(k0 ). 6 7 4 u(k0 þ 1) 5 u(k0 ) AB A2 B ÁÁÁ ANÀ1 2 (4:123) Â B AB A2 B ÁÁÁ ANÀ1 (4:124) Â Equation 4. THEOREM 4. where Â Ã Q ¼ B AB A2 B Á Á Á ANÀ1 B (4:121) Proof: We assume that the system has one input.204 Control of Color Imaging Systems: Analysis and Design 4.124 has a solution if matrix Q ¼ B rank. then the state of the system at time k ¼ N þ k0 is given by x(N þ k0 ) ¼ AN x(k0 ) þ Nþk0 À1 X n¼k0 ANþk0 À1Àn Bu(n) (4:122) To drive the system to origin.14 Consider the following system: x(k þ 1) ¼ 1 1:5 Ã B is full ! ! À0:25 À1 x(k) þ u(k) À0:25 1 .. u(k0 þ 1).

a system is completely state observable if the states of the system can be found from the knowledge of inputs and outputs of the system. Q is not full rank and the system is not completely state controllable.15 Consider the SISO system 0:5 1 x(k þ 1) ¼ 4 0 0:75 0 0 The controllability matrix Q is Â 1 A B ¼ 4 À1 2 2 2 3 2 3 0 1 1 5x(k) þ 4 À1 5u(k) 0:8 2 Q¼ B AB Ã 2 3 À0:5 1 1:25 2:54 5 1:6 1:28 Since det (Q) ¼ À9:74 6¼ 0. Q is full rank and the system is completely state controllable.State-Variable Representation The controllability matrix Q is Q ¼ ½B AB ¼ À1 1 À1:25 À1:75 ! 205 Since det (Q) ¼ 3 6¼ 0.16 Consider the MIMO system 1 x(k þ 1) ¼ 4 0 0 The controllability matrix Q is Â 1 0 A2 B ¼ 4 0 1 À1 2 Ã 2 1 0 0 1 À1 2 1 0 À1 3 0 15 2 2 3 2 3 ! 0 0 1 0 u (k) 1 0 5x(k) þ 4 0 1 5 1 u2 (k) 0 1 À1 2 Q¼ B AB Since rank(Q) ¼ 2. Q is full rank and the system is completely state controllable. Example 4. 4. Since we know the . Example 4.10 OBSERVABILITY OF LTI SYSTEMS In general.

5 . THEOREM 4. Since the results obtained are identical to continuous and discrete-time systems. 7 4 . In this section.10. we examine the observability of LTI system. where Q is deﬁned as 2 C 3 (4:126) 6 CA 7 6 7 6 7 6 CA2 7 P¼6 7 6 . then kÀ1 X n¼0 y(k) ¼ Cx(k) þ Du(k) ¼ CAk x(0) þ C AkÀ1Àn Bu(n) þ Du(k) (4:127) . CANÀ1 Proof: Let the initial state be x(0) at time k ¼ 0 and the input and output for time k ! 0 be u(k) and y(k).10.1 DEFINITION OF OBSERVABILITY A discrete-time LTI system is said to be completely state observable if the states of the system can be estimated from the knowledge of inputs and outputs of the system. 4.206 Control of Color Imaging Systems: Analysis and Design system and the inputs to the system.7].2 A discrete-time LTI system given by the input and output equations x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) þ Du(k) (4:125) is completely state observable if the observability matrix Q is full rank. respectively.2 OBSERVABILITY CONDITION The following theorem gives the necessary and sufﬁcient condition for observability of a discrete-time linear system. then the system is completely state observable. we only consider observability of discrete-time LTI system [6. Therefore. 4. if we can ﬁnd all the initial conditions from inputs and outputs of the system. the only unknown in determining the states of the system is the initial conditions.

4 NÀ2 P NÀ2Àn 5 A B CANÀ1 u(N À 1) y(N À 1) 2 C CA CA2 . 3 2 C CA CA2 . . 6 7 6 7 . . the states of the system at any other time can be obtained by solving the state equations. 3 n¼0 2 (4:129) Equation 4. 3 n¼0 2 (4:128) or 3 2 3 0 u(0) AB 6 7 6 7 6 u(1) 7 6 7 6 7 6 7 6 7 6 7 6 AB þ B 7 7 6 7x(0) ¼ 6 6 7 À D6 u(2) 7 . Example 4. . . 4 NÀ2 5 P NÀ2Àn A B CANÀ1 u(N À 1) y(N À 1) 2 y(0) y(1) y(2) . 4 5 4 5 .18 Consider the SISO system 0:5 1 x(k þ 1) ¼ 4 1 0:75 2 À1 y(k) ¼ ½ 1 2 3 2 3 0 1 1 5x(k) þ 4 À1 5u(k) 0:8 2 0 À2 x(k) þ 3u(k) . 6 7 6 7 4 . 6 7 6 7¼6 7 6 7 . 3 2 y(0) y(1) y(2) . . 5 6 7 . 4 5 6 7 . .17 Consider the following system x(k þ 1) ¼ 1 1:5 ! ! À1 À0:25 u(k) x(k) þ 1 À0:25 y(k) ¼ ½ 1 0 x(k) The observability matrix P is P¼ ! ! 1 0 C ¼ 1 À0:25 CA Since det (P) ¼ À0:25 6¼ 0.127 can be written in matrix form as 3 2 3 0 u(0) AB 6 7 6 u(1) 7 6 7 6 7 6 7 6 7 6 7 6 7 6 AB þ B 7 6 7 6 7x(0) þ C 6 7 þ D6 u(2) 7 . . 4 5 4 5 .129 can be used to solve for x(0) if and only if matrix P is full rank. Once the initial state is obtained. P is full rank and the system is completely state observable. Example 4. . . 6 7 À C6 6 7 6 7 7 .State-Variable Representation 207 Equation 4. .

P is not full rank and the system is not completely state observable. Example 4. Example 4.208 Control of Color Imaging Systems: Analysis and Design The observability matrix P is 3 2 1 C P ¼ 4 CA 5 ¼ 4 À3:5 À1:95 CA2 2 3 0 À2 3 À1:6 5 0:35 1:72 Since det (P) ¼ À3:53 6¼ 0. P is full rank and the system is completely state observable.20 Consider the MIMO system 2 3 3 ! 1 0 1 À1 0 u (k) x(k þ 1) ¼ 4 0 0:5 1 5x(k) þ 4 À3 1 5 1 u2 (k) À1 0 À1 0 0:6 ! À1 1 1 x(k) y(k) ¼ 1 1 0 2 .19 Consider the MIMO system 2 1 0 0 3 2 1 3 0 " # 7 u1 (k) 17 5 u2 (k) 2 6 x(k þ 1) ¼ 6 0 1 4 " y(k) ¼ 0 0 À1 0 1 1 7 6 0 7x(k) þ 6 0 5 4 1 1 0 # x(k) À1 The observability matrix P is À1 3 6 1 2 6 C 6 7 6 À1 6 P ¼ 4 CA 5 ¼ 6 6 1 6 CA2 6 4 À1 1 2 3 0 1 1 07 7 7 0 17 7 1 07 7 7 0 15 1 0 Since rank(P) ¼ 2.

3 Obtain the response y(t) of the following system.2 Find the state-space representation of the following dynamical systems: (a) d3 y(t) d2 y(t) dy(t) du(t) þ 2y(t) ¼ þ 8u(t) þ3 þ6 dt 3 dt 2 dt dt (b) y(k þ 3) þ 0:5y(k þ 2) þ y(k þ 1) þ 0:89y(k) ¼ 2:4u(k) 4. when u(k) ¼ (À1)k for k ! 0: x1 (k þ 1) x2 (k þ 1) ! ¼ À1 1 ! 1 0 À1 x1 (k) x2 (k) À0:25 0 ! x1 (k) x2 (k) ! ! þ 1 0 ! u(k) x(0) ¼ y(k) ¼ ½ 1 .1 Find the state-space representation of the following dynamical systems in controllable and observable canonical forms: (a) d3 y(t) d2 y(t) dy(t) þ2 þ3 þ y(t) ¼ 4u(t) 3 dt dt 2 dt (b) y(k þ 2) þ y(k þ 1) þ 0:8y(k) ¼ 8u(k þ 1) þ 6u(k) 4.4 Obtain the response y(k) of the following system. P is full rank and the system is completely state observable. PROBLEMS 4. where u(t) is the unit step function: ! ! ! ! _ À1 À0:75 x1 (t) 0 x1 (t) ¼ þ u(t) _ 1 0 1 x2 (t) x2 (t) ! 1 x(0) ¼ 0 ! x1 (t) y(t) ¼ ½ 1 0 x2 (t) 4.State-Variable Representation The observability matrix P is 209 À1 6 3 6 1 2 6 C 6 7 6 À2 6 P ¼ 4 CA 5 ¼ 6 6 1 6 CA2 6 4 À3:6 0 2 1 1 1:5 À0:5 2:75 À1:25 7 0 7 7 1:6 7 7 7 1 7 7 7 2:46 5 0:1 1 3 Since rank(P) ¼ 3.

7 The state-transition matrix of a time-varying system is the solution to the following partial differential equation: @w(t. t). with boundary condition w(t. t) ¼ I @t Find the state-transition matrix of the following systems: ! ! _ eÀt 0 x1 (t) ¼ (a) _ x2 (t) 0 À2t ! ! _ À1 t x (t) ¼ (b) 1 _ x2 (t) 0 À1 ! x1 (t) x2 (t) ! x1 (t) x2 (t) 4.210 Control of Color Imaging Systems: Analysis and Design 4. t) ¼ A(t)w(t.6 Show that the discrete-time state-transition matrix w(k) ¼ Ak satisﬁes the following properties: (a) w(0) ¼ I (b) w(Àk) ¼ AÀk ¼ (Ak )À1 ¼ wÀ1 (k) (c) w(k1 þ k2 ) ¼ w(k1 )w(k2 ) ¼ w(k2 )w(k1 ) (d) w(k3 À k2 )w(k2 À k1 ) ¼ w(k3 À k1 ) 4.8 An LTI system is described by the following differential equations: ! _ À6 x1 (t) ¼ _ À2 x2 (t) 4 0 ! ! ! x1 (t) 0 þ u(t) 1 x2 (t) .5 Show that the continuous-time state-transition matrix w(t) ¼ eAt satisﬁes the following properties: (a) w(0) ¼ I (b) w(Àt) ¼ eÀAt ¼ (eAt )À1 ¼ wÀ1 (t) (c) w(t1 þ t2 ) ¼ w(t1 )w(t2 ) ¼ w(t2 )w(t1 ) (d) w(t3 À t2 )w(t2 À t1 ) ¼ w(t3 À t1 ) 4.

(b) Solve the decoupled system for the new variables z1 (t) and z2 (t). and demonstrate that your results are consistent with the results obtained for x1 (k) and x2 (k).12 Consider the system deﬁned by ! _ 1 x1 (t) ¼ _ À3 x2 (t) 2 ! x1 (t) ! þ À6 1 ! u(t) À5 x2 (t) ! x1 (t) y(t) ¼ ½ 2 3 x2 (t) (a) Transform the system equations into the controllable canonical form. (b) Solve the decoupled system for the new variables z1 (k) and z2 (k). 4. 4.9 A discrete LTI system is described by the following state equations: ! 0 x1 (k þ 1) ¼ À0:5 x2 (k þ 1) 1 1:5 ! ! ! x1 (k) À1 þ u(k) 1 x2 (k) Let x1 (0) ¼ 1.11 Consider the following transfer function H(z) ¼ 3z2 þ 2z þ 1 z2 þ 0:7z þ 0:12 Obtain the state-space representation of this system in controllable and observable canonical forms. x2 (0) ¼ À1.10 Consider the continuous transfer function H(s) ¼ 3s þ 4 s2 þ 6s þ 8 Obtain the state-space representation of this system in controllable and observable canonical forms. and demonstrate that your results are consistent with the results obtained for x1 (t) and x2 (t). . (a) Find a nonsingular transformation matrix T such that the change in variables z(k) ¼ Tx(k) will lead to a decoupled system of equations. and u(t) ¼ eÀt . derive expressions for x1 (t) and x2 (t). and u(k) ¼ (0:5)k . x2 (0) ¼ À1. 4. (a) Find a nonsingular transformation matrix T such that the change in variables z ¼ Tx will lead to a decoupled system of differential equations. (b) Transform the system equations into the observable canonical form. derive expressions for x1 (k) and x2 (k). 4.State-Variable Representation 211 Let x1 (0) ¼ 1.

ﬁnd the uncontrollable= unobservable modes. 1 0 0 1 3 2 7 6 0 5 x(k) þ 4 b21 b31 b11 b12 b22 b32 b13 b33 3 7 b23 5 u(k) . (b) Find the necessary and sufﬁcient condition for controllability of this system.14 Is the following system completely state controllable and observable? 2 6 7 6 4 x2 (k þ 1) 5 ¼ 4 0 x3 (k þ 1) y(k) ¼ ½ 2 À3 À3 x1 (k þ 1) 3 2 0 1 0 2 3 0 76 7 6 7 1 54 x2 (k) 5 þ 4 0 5 u(k) 0 x1 (k) x3 (k) 1 32 3 À1 À2 2 3 x1 (k) 6 7 À2 4 x2 (k) 5 x3 (k) 4. 4.212 Control of Color Imaging Systems: Analysis and Design 4. (b) Is the system controllable? Is it observable? If not.13 Consider the system deﬁned by 2 7 6 6 4 x2 (k þ 1) 5 ¼ 4 0 x3 (k þ 1) " y(k) ¼ 0:3 1 0 0 1 x1 (k þ 1) 3 2 0 1 0 0:4 0 32 76 7 6 1 54 x2 (k) 5 þ 4 1 0 x1 (k) 3 2 0 3 # 1 " 7 u1 (k) 05 u2 (k) 1 1 x3 (k) 2 3 # x1 (k) 0 6 7 4 x2 (k) 5 0 x3 (k) (a) Is the system completely state controllable? (b) Is the system completely state observable? (c) Is the system completely output controllable? 4.16 Consider the system given by 2 6 x(k þ 1) ¼ 4 0 1 0 0 y(k) ¼ x1 (k) Y(z) (a) Find the transfer function U(z).15 For the system 1 x(k) x(k þ 1) ¼ y(k) ¼ ½ 1 0 À0:125 1 0:75 ! x(k) þ 0 2 ! u(k) and Y(z) (a) Find the transfer function H(z) ¼ U(z).

3. 2001. 7. Tewari. 2005.17 Show that the system given by x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) where x 2 RN . 1985. Ogata. Discrete-Time Control Systems. McGrawHill. REFERENCES 1. NJ. D. Powell. 1995. A. J. Kuo. Prentice-Hall. 6. L. and y 2 RP is completely output controllable if and only if the controllability matrix P given as P ¼ ½ CB CAB: CA2 B Á Á Á CANÀ1 B is of rank M. 1990. 4. Franklin. Ogata. Upper Saddle River. Control System Design: An Introduction to State-Space Methods. Emami-Naeini.State-Variable Representation 213 4. and A. Friedland.C. W. Brogan. Englewood. Oxford University Press. NJ. New York. K. Englewood. Prentice Hall. Digital Control Systems. NJ. 2. u 2 RM . Upper Saddle River. Modern Control Engineering. Wiley. . Modern Control Design with MATLAB and Simulink. Prentice Hall. Feedback Control of Dynamic Systems. NJ. 2002. K. New York. Prentice Hall. Modern Control Theory. B. B. G. 5. United Kingdom. Chichester. 1994.

.

the closed-loop control system will be described by the following equation: x(k þ 1) ¼ (A À BK)x(k) (5:3) The block diagram of the closed-loop system with state feedback appears in Figure 5. u(k) ¼ ÀKx(k) (5:2) (5:1) Then. 215 . The state estimation is another topic that will be discussed later in this chapter. the state feedback control signal is generated by using weighted sum of the states. We conclude this chapter by introducing optimal control for design of closed-loop control systems.2 STATE FEEDBACK 5.1 INTRODUCTION This chapter covers the design of closed-loop control systems in state space.2. The feedback gain is designed through pole placement.1. This approach is valid when all the states are available.1 BASIC CONCEPT Consider a single-input single-output (SISO) open-loop LTI discrete-time system in state-space form given by x(k þ 1) ¼ Ax(k) þ Bu(k) where A 2 RNÂN B 2 RN u 2 R1 x 2 RN The block diagram of the open-loop system is shown in Figure 5. In other words. where the poles of the closed-loop system are assigned for a speciﬁc time response.5 Closed-Loop System Analysis and Design 5. The well-known linear quadratic regulator (LQR) will be used as an example for design of optimal control loops.2. 5. For an N-state SISO system. where the control signal is made to be proportional to the states of the system. The most common technique is state feedback. In practical applications not all states are measurable and they have to be estimated from control input and measured system output. assuming states are accessible.

the solution to the closed-loop system of Equation 5. The following example illustrates this concept. x(k + 1) u(k) B + + Unit delay x(k) A –K FIGURE 5. Example 5.216 Control of Color Imaging Systems: Analysis and Design x(k + 1) u(k) B + + Unit delay x(k) A FIGURE 5. The feedback gain K is a vector for SISO system.1 Block diagram of the open-loop control system.1 Consider an open-loop control system given by the following state equation: x(k þ 1) ¼ 0 À0:72 ! ! 0 1 u(k) x(k) þ 1 1:7 (5:5) . which can be designed to place the eigenvalues of the closed-loop system inside the unit circle even if the original open-loop system is unstable or does not give an acceptable transient response.3 is x(k) ¼ (A À BK)k x(0) (5:4) Therefore. the states of the system will approach zero for any arbitrary initial conditions if the eigenvalues of the closed-loop matrix Ac ¼ A À BK are inside the unit circle in the complex z-plane.2 Block diagram of the closed-loop system with state feedback. Using techniques described in Chapter 4.

the two states of the system are x1 (k) ¼ À11(0:8)k þ 12(0:9)k and x2 (k) ¼ À9(0:8)k þ 11(0:9)k (5:8) (5:7) Plot of x1 (k) is shown in Figure 5.3. the response of the system is slow and it takes more than 50 time steps to converge to the origin. As is seen in this example. . In order to improve the transient behavior of the system.5 1 0.3 Transient behavior of the open-loop system.5 x1 (k) 2 1.5 3 2. The zero-input response of the system for the initial condition of x(0) ¼ ½ 1 À2 T is given by x(k) ¼ 0 À0:72 1 1:7 !k x(0) ¼ 9(0:8)k À 8(0:9)k 7(0:8)k À 7(0:9)k À10(0:8)k þ 10(0:9)k À8(0:8)k þ 9(0:9)k ! 1 À2 ! (5:6) Therefore. The closed-loop system matrix Ac is given by Ac ¼ A À BK ¼ 0 À0:72 ! ! 0 1 À ½ k1 1 1:7 k2 ¼ 0 À0:72 À k1 1 1:7 À k2 ! (5:9) 4 3.Closed-Loop System Analysis and Design 217 The poles of the open-loop system or the eigenvalues of the A matrix are l1 ¼ 0:9 and l2 ¼ 0:8. we use state feedback.5 0 0 5 10 15 20 25 k 30 35 40 45 50 FIGURE 5.

As it can be seen. the closed-loop response converges much faster than the openloop response. In the next section. Clearly. Location of poles directly affects the state response. which is one of the reasons why closed-loop systems are preferred over openloop design. we select the closedloop poles to be at l1 ¼ l2 ¼ 0:2 (5:11) This means that the characteristic polynomial of the closed-loop system should be P(l) ¼ (l À 0:2)(l À 0:2) ¼ l2 À 0:4l þ 0:04 Comparing the two equations (Equations 5. we have & 0:72 þ k1 ¼ 0:04 k2 À 1:7 ¼ À0:40 (5:13) (5:12) Therefore & k1 ¼ À0:68 k2 ¼ 1:3 (5:14) The closed-loop system response is given by ! 0 1 k x(k) ¼ (A À BK)k ¼ x(0) À0:04 0:4 " # ! 1 5k(0:2)k (1 À k)(0:2)n ¼ À0:2k(0:2)n (1 þ k)(0:2)k À2 Therefore the two states of the closed-loop system are x1 (k) ¼ (1 À 11k)(0:2)k and x2 (k) ¼ À(2 þ 2:2k)(0:2)k (5:17) (5:16) (5:15) The response of the ﬁrst state of the open-loop and closed-loop systems is shown in Figure 5.218 Control of Color Imaging Systems: Analysis and Design The characteristic polynomial of the closed-loop system is P(l) ¼ jlI À Ac j ¼ l(l À 1:7 þ k2 ) þ 0:72 þ k1 ¼ l2 þ (k2 À 1:7)l þ 0:72 þ k1 (5:10) To improve the transient response of the closed-loop system. the state feedback has led to a controllable response to the state. .10 and 5.12).4. we discuss the pole-placement design for SISO systems using state feedback [1–3].

2 POLE-PLACEMENT DESIGN OF SISO SYSTEMS (5:18) Consider the SISO dynamic system described by the following state equation: x(k þ 1) ¼ Ax(k) þ Bu(k) With the assumption that the system is completely state controllable using the state feedback u(k) ¼ ÀKx(k). . 5.Closed-Loop System Analysis and Design 4 219 3 Open-loop response 2 x1(k) 1 0 Closed-loop response –1 –2 0 5 10 15 20 25 k 30 35 40 45 50 FIGURE 5. lN . lN . . This means that the eigenvalues of the closed-loop matrix Ac ¼ A À BK are at the desired locations given by l1 .4 Open-loop and closed-loop response. we have AN þ a1 ANÀ1 þ a2 ANÀ2 þ Á Á Á þ aN I ¼ P(Ac ) ¼ 0 c c c But P(Ac ) ¼ aN I þ aNÀ1 Ac þ aNÀ2 A2 þ Á Á Á þ AN c c ¼ aN I þ aNÀ1 (A À BK) þ aNÀ2 (A À BK)2 þ Á Á Á þ (A À BK)N (5:21) (5:20) (5:19) . Therefore the characteristic polynomial of Ac is P(l) ¼ jlI À Ac j ¼ (l À l1 )(l À l2 ) Á Á Á (l À lN ) ¼ lN þ a1 lNÀ1 þ a2 lNÀ2 þ Á Á Á þ aN By Cayley–Hamilton theorem (Section 3. we wish to place the poles of the closed-loop system at l1 . . .3).11. . . l2 . . . l2 .2.

therefore we have 2 3 aNÀ1 K þ aNÀ2 KAc þ Á Á Á KANÀ1 c 6 aNÀ2 K þ aNÀ3 KAc þ Á Á Á KANÀ2 7 c 6 7 6 7 Â 6 aNÀ3 K þ aNÀ4 KAc þ Á Á Á KANÀ3 7 ¼ B c 6 7 6 7 . the controllability matrix Â Ã B AB A2 B Á Á Á ANÀ1 B is nonsingular and its inverse exists. 4 5 . 4 5 . .23 can be written in matrix form as 3 aNÀ1 K þ aNÀ2 KAc þ Á Á Á KANÀ1 c 6 aNÀ2 K þ aNÀ3 KAc þ Á Á Á KANÀ2 7 c 6 7 7 Ã6 NÀ1 6 aNÀ3 K þ aNÀ4 KAc þ Á Á Á KANÀ3 7 c A B 6 7 6 7 . K AB A2 B Á Á Á ANÀ1 B ÃÀ1 P(A) (5:26) .21 can be rewritten as P(Ac ) ¼ aN I þ aNÀ1 (A À BK) þ aNÀ2 (A2 À ABK À BKAc ) þ aNÀ3 (A3 À A2 BK À ABKAc À BKA2 ) c þ Á Á Á þ (AN À ANÀ1 BK À ANÀ2 BKAc À Á Á Á À BKANÀ1 ) c Simplifying further. 4 5 . . we have 3 aNÀ1 K þ aNÀ2 KAc þ Á Á Á KANÀ1 c 6 aNÀ2 K þ aNÀ3 KAc þ Á Á Á KANÀ2 7 c 6 7 7 Ã6 NÀ1 6 aNÀ3 K þ aNÀ4 KAc þ Á Á Á KANÀ3 7 c ÁÁÁ A B 6 7 6 7 . . K (5:24) 2 2 (5:23) (5:22) Â P(Ac ) ¼ P(A) À B AB A2 B ÁÁÁ Since P(Ac ) ¼ 0. K (5:25) Â P(A) ¼ B AB A2 B Since the system is completely state controllable.220 Control of Color Imaging Systems: Analysis and Design Equation 5. this results in P(Ac ) ¼ aN I þ aNÀ1 A þ aNÀ2 A2 þ Á Á Á þ AN À aNÀ1 BK À aNÀ2 ABK À aNÀ2 BKAc À Á Á Á À BKANÀ1 c ¼ P(A) À B(aNÀ1 K þ aNÀ2 KAc þ Á Á Á þ KANÀ1 ) c À AB(aNÀ2 K þ aNÀ3 KAc þ Á Á Á þ KANÀ2 ) À Á Á Á À ANÀ1 BK c Equation 5.

4]. The second algorithm is summarized below.26 can be used to solve for K by premultiplying both sides of this equation by ½ 0 0 Á Á Á 0 1 . Example 5.27 K ¼ ½0 ¼ ½0 AB À1 P(A) ¼ ½ 0 1 ½ B AB À1 (A2 À 0:4A þ 0:04I) ! ! 0 1 À1 À0:68 1:3 1 ¼ ½ À0:68 1:3 À0:93 1:53 1 1:7 1 ½ B There are other algorithms that can be used for pole placement [1.Closed-Loop System Analysis and Design 221 Equation 5.2 Consider the dynamic system given by 0 x(k þ 1) ¼ À0:72 ! ! 0 1 u(k) x(k) þ 1 1:7 Design a state feedback to place the closed-loop poles at l1 ¼ l2 ¼ 0:2 SOLUTION The characteristic polynomial of the closed-loop system is P(l) ¼ (l À l1 )(l À l2 ) ¼ (l À 0:2)(l À 0:2) ¼ l2 À 0:4l þ 0:04 The controllability matrix is ½B AB ¼ 0 1 1 1:7 ! Since the controllability matrix is full rank. For the desired poles. a. The resulting gain matrix K is given as K ¼ ½0 0 ÁÁÁ Â 0 1 B AB A2 B Á Á Á ANÀ1 B ÃÀ1 P(A) (5:27) The above equation is known as Ackermann’s formula for pole placement. ﬁnd the desired characteristic polynomial of the closed-loop system Pc (l) ¼ jlI À Ac j ¼ (l À l1 )(l À l2 ) Á Á Á (l À lN ) ¼ lN þ a1 lNÀ1 þ a2 lNÀ2 þ Á Á Á þ aN (5:28) . The feedback gain vector is computed using Equation 5. the system is completely state controllable and pole placement is possible.

6 . . Find the gain matrix K using the following equation: K ¼ ½ aN À bN aNÀ1 À bNÀ1 ÁÁÁ a1 À b1 T À1 (5:32) Example 5. The transformation T is given by T ¼ QW ¼ ½ B AB b1 1 ! ! 1 1 1 1 ¼ 1 À1:12 1 0 ! 2 1 ¼ À0:12 0 1 1 ! .222 Control of Color Imaging Systems: Analysis and Design b.3 Consider the dynamic system given by x(k þ 1) ¼ 0 À0:12 ! ! 1 1 u(k) x(k) þ 1 À1 Design a state feedback controller to place the closed-loop poles at l1 ¼ 0:3 À j0:4 and l2 ¼ 0:3 þ j0:4 SOLUTION The characteristic polynomial of the open-loop system is l À1 2 P(l) ¼ jlI À Aj ¼ 0:12 l þ 1 ¼ l þ l þ 0:12 (5:33) Comparing Equation 5. . . we have b1 ¼ 1. The characteristic polynomial of the desired closed-loop system is Pc (l) ¼ (l À 0:3 þ j0:4)(l À 0:3 þ j0:4) ¼ l2 À 0:6l þ 0:25 Hence a1 ¼ À0:6. and a2 ¼ 0:25. Find the characteristic polynomial of the open-loop system P(l) ¼ jlI À Aj ¼ lN þ b1 lNÀ1 þ b2 lNÀ2 þ Á Á Á þ bN c. 4 b 1 1 2 bNÀ2 bNÀ1 .29.7 05 0 (5:30) (5:29) (5:31) d.33 with Equation 5. and b2 ¼ 0:12.7 . W ¼6 . Find the transformation T given by T ¼ QW where Q is the controllability matrix and W is given by bNÀ1 6 bNÀ1 6 6 . . 1 0 ÁÁÁ ÁÁÁ ÁÁÁ ÁÁÁ b1 1 . 0 0 3 1 07 7 .7 .

4 Consider the dynamic system given by 0 x(k þ 1) ¼ À0:12 ! ! 1 0 x(k) þ u(k) À1 1 Design a state feedback to place the closed-loop poles at l1 ¼ l2 ¼ 0 SOLUTION The characteristic polynomial of the open-loop system is P(l) ¼ l2 þ l þ 0:12 Hence b1 ¼ 1. and b2 ¼ 0:12. The characteristic polynomial of the desired closed-loop system is Pc (l) ¼ l2 Hence.Closed-Loop System Analysis and Design The gain matrix K is given as K ¼ ½ a2 À b2 a1 À b1 T À1 ¼ ½ 0:13 ¼ ½ À0:0292 À1:5708 À1:6 2 À0:12 1 1 !À1 223 Example 5. a1 ¼ a2 ¼ 0 and the transformation T is given by T ¼ QW ¼ ½ B AB The gain matrix K is given as K ¼ ½ a2 À b2 a1 À b1 T À1 b1 1 ! ! ! ! 1 0 1 1 0 1 1 ¼ ¼ 0 1 1 À1 1 0 0 ¼ ½ À0:12 1 0 À1 0 1 !À1 ¼ ½ À0:12 À1 The closed-loop state equations are x(k þ 1) ¼ 0 À0:12 0 ¼ À0:12 ! 1 x(k) À À1 ! 1 x(k) À À1 ! 0 Kx(k) 1 ! 0 0 ½ À0:12 À1 x(k) ¼ 1 0 ! 1 x(k) 0 .

. . KM2 ÁÁÁ ÁÁÁ ÁÁÁ K1N K2N . These unknowns are entries of the gain matrix K.2. . 4 . 5. KM1 is given as K12 K22 . Therefore there are fewer equations than number of unknowns which means that there are inﬁnite number of possible solutions.37 and 5. the closed-loop characteristic polynomial in terms of the desired poles is given by Pc (l) ¼ (l À l1 )(l À l2 ) Á Á Á (l À lN ) ¼ lN þ a1 lNÀ1 þ a2 lNÀ2 þ Á Á Á þ aN (5:38) Comparing Equations 5. then Now consider an arbitrary initial state x(0) ¼ ½ x1 (0) 0 x(1) ¼ 0 ! 0 1 x(0) ¼ 0 0 1 0 ! ! ! x1 (0) 0 ¼ 0 x2 (0) Therefore any arbitrary initial state is driven to zero in one step. This is called dead beat control. The MATLAB1 function ﬁle named .3 POLE-PLACEMENT DESIGN (MIMO) SYSTEMS OF MULTIPLE-INPUT MULTIPLE-OUTPUT Consider an open-loop MIMO system in state-space form given by x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) where A 2 RNÂN B 2 RNÂM u 2 RM C 2 RPÂN y 2 RP The state feedback control law is given by u(k) ¼ ÀKx(k) where the gain matrix K is M Â N and 2 K11 6 K21 6 K¼6 . we get N equations with M Â N unknowns.224 Control of Color Imaging Systems: Analysis and Design x2 (0) T . . .38. KMN 3 7 7 7 5 (5:36) (5:35) (5:34) The characteristic polynomial of the closed-loop system is Pc (l) ¼ jlI À A þ BK j (5:37) On the other hand.

For example if we choose K11 ¼ K22 ¼ 0. Example 5. we have & K11 þ K22 À K21 þ 1 ¼ À0:5 K11 (1 þ K22 ) þ 0:12(1 þ K22 ) À K12 (0:12 þ K21 ) ¼ 0:06 As can be seen. and l2 ¼ 0:2.Closed-Loop System Analysis and Design 225 ‘‘place’’ implements the Ackermann algorithm for SISO systems. we have & ÀK21 þ 1 ¼ À0:5 0:12 À K12 (0:12 þ K21 ) ¼ 0:06 The solutions to the above equations are & K21 ¼ 1:5 K12 ¼ 0:037 K12 K22 ! .P). where A and B are system matrices and P is a vector containing desired poles of the closed-loop system.B. For MIMO systems. we have two equations and four unknowns. the ‘‘place’’ algorithm uses extra degrees of freedom to obtain a robust solution for matrix K. It minimizes the sensitivity of the closed-loop poles to the variations in system’s parameters A and B. then Pc (l) ¼ (l À 0:3)(l À 0:2) ¼ l2 À 0:5l þ 0:06 Comparing the two equations of the closed-loop characteristic polynomials. Therefore the solution is not unique.5 Consider the MIMO dynamic system given by ! 1 0 1 x(k) þ x(k þ 1) ¼ 0 À0:12 À1 ! ! À1 1 u1 (k) u2 (k) Design a state feedback to place the closed-loop poles at l1 ¼ 0:3 l2 ¼ 0:2 SOLUTION The system matrix of the closed-loop system is ! ! 1 À1 K11 0 1 À Ac ¼ A À BK ¼ 0 1 À0:12 À1 K21 ! ÀK11 þ K21 1 À K12 þ K22 ¼ À0:12 À K21 À1 À K22 The characteristic polynomial of the closed-loop system is Pc (l) ¼ jlI À Ac j ¼ l2 þ (K11 þ K22 À K21 þ 1)l þ K11 (1 þ K22 ) þ 0:12(1 þ K22 ) À K12 (0:12 þ K21 ) Since the desired closed-loop poles are at l1 ¼ 0:3. The command to implement pole placement in MATLAB is K ¼ place(A.

For example consider slight perturbation of matrix A by DA. the second design using the MATLAB pole-placement algorithm is less sensitive to variations in system’s parameters.40) is AC ¼ ! ! 1 À1 À0:32 0:0039 0:9996 À À0:12 0 1 0:1174 À1:0048 ! 0:2039 À0:0004 ¼ 0:0026 0:2952 À0:3 À1:3 ! The closed-loop poles are p2 ¼ ½ 0:2952 0:2039 T and the percentage change is Dl ¼ k p À p2k2 Â 100 ¼ 1:72% k p k2 Therefore.226 Control of Color Imaging Systems: Analysis and Design The resulting gain matrix is K¼ K11 K21 K12 K22 ! ¼ 0 1:5 0:037 0 ! (5:39) The MATLAB pole placement yields K¼ K11 K21 K12 K22 ! ¼ À0:32 À0:12 À0:3 À1:3 ! (5:40) The MATLAB solution gives a more robust design with respect to variations in the system parameters. .39) is AC ¼ A À BK ¼ ¼ 1:5039 1:6174 ! 0:0039 0:9996 1 À 0:1174 À1:0048 0 ! 0:9626 À1:0048 À1 1 ! 0 0:037 1:5 0 ! The closed-loop poles are eigenvalues of AC which are p1 ¼ ½ 0:3784 0:1207 T . The percentage change from the desired poles p ¼ ½ l1 l2 T ¼ ½ 0:3 0:2 T is Dl ¼ k p À p1k2 Â 100 ¼ 30:92% k p k2 The closed-loop A matrix using the second design (K matrix given by Equation 5. A ¼ A þ DA ¼ 0 À0:12 ! ! ! 0:0039 0:9996 0:0039 À0:0004 1 ¼ þ 0:1174 À1:0048 0:0026 À0:0048 À1 The closed-loop A matrix using the ﬁrst design (K matrix given by Equation 5. that is.

The response of the system for different values of l1 . Let the plant to be controlled be given in state space as x(k þ 1) ¼ Ax(k) þ Bu(k) (5:46) . we consider a second-order dynamic system with zero input. l2 are eigenvalues (poles) of the closed-loop dynamic system which can be assigned arbitrarily.4 RELATIONSHIP BETWEEN POLES AND THE CLOSED-LOOP SYSTEM RESPONSE The closed-loop system response is directly related to the locations of the poles in the complex plane. The time response of the system for the given initial condition is given by x(k) ¼ Ak x(0) ¼ This can be simpliﬁed as x1 (k) ¼ [Àlk þ 2lk ]u(k) 1 2 and x2 (k) ¼ [À3lk þ 5lk ]u(k) 1 2 (5:45) (5:44) À5l1 þ 6l2 À15l1 þ 15l2 2l1 À 2l2 6l1 À 5l2 !k 1 2 ! (5:43) For the system to be stable the poles are chosen to be inside the unit circle in complex plane. and jl2 j < 1.3 LQR DESIGN 5.2. In the derivation of the LQR.1. Assume that the system is initially at x1 (0) ¼ 1 and x2 (0) ¼ 2. This means that jl1 j < 1.Closed-Loop System Analysis and Design 227 5. and l2 are shown in Table 5. To illustrate the relationship between the poles and the system response.3. we assume that all the states of the system are available to the controller. 5. Assume that the closed-loop second-order system is given by x(k þ 1) ¼ Ax(k) where matrix A is given in the diagonalized form as A¼ 1 3 À2 À5 ! l1 0 0 l2 ! 1 3 À2 À5 !À1 ¼ À5l1 þ 6l2 À15l1 þ 15l2 2l1 À 2l2 6l1 À 5l2 ! (5:42) (5:41) Here l1 . It provides an optimal design based on a speciﬁed ﬁgure of merit which is quadratic in both states as well as in control.5].1 INTRODUCTION The LQR is a well-known optimal control design technique that results in a feedback gain similar to the state feedback design [4.

4 0.6 Time Response Response Type Underdamped x1(k) 0.8 0.5 0.3 0.8 0.9 0.7 0.5 0.2 0.3 0.4 0.7 0.1 System Time Response and Pole Locations Poles l1 ¼ 0:25 l2 ¼ 0:25 1 0.5 1 Oscillation x1(k) 0 –0.1 0 0 2 4 6 8 10 12 14 16 18 20 k l1 ¼ À0:42 l2 ¼ À0:42 0.9 0.5 0 1 2 3 4 5 6 7 8 9 10 k .6 Deadbeat x1(k) 0.228 Control of Color Imaging Systems: Analysis and Design TABLE 5.1 0 0 1 2 3 4 5 6 7 8 9 10 k l1 ¼ 0 l2 ¼ 0 1 0.2 0.

.3. 5. The goal is to minimize performance index J with respect to the control signal u(k) subject to the constraints given by the state equation of the plant (Equation 5. The closed-loop system dynamic is governed by the state equation given as x(k þ 1) ¼ (A À BK(k))x(k) Let us form a quadratic performance function as follows J ¼ 1 xT (N)Sx(N) þ 1 2 2 NÀ1 P k¼0 (5:48) xT (k)Qx(k) þ uT (k)Ru(k) (5:49) where S and Q are positive or positive semideﬁnite matrices R is a positive deﬁnite matrix The ﬁrst term in the performance index penalizes the deviation of the ﬁnal state from the origin. we form a new performance index H as follows H ¼ 1 xT (N)Sx(N) þ 1 2 2 NÀ1 P k¼0 (5:51) xT (k)Qx(k) þ uT (k)Ru(k) (5:52) þ lT (k þ 1)½ x(k þ 1) À Ax(k) À Bu(k) where l(k) ¼ ½ l1 (k) l2 (k) Á Á Á lN (k) T . The term xT (k)Qx(k) penalizes deviation of the states from the origin during the control process and the term uT (k)Qu(k) is a measure of energy used by the control signal.2 SOLUTION OF THE LQR PROBLEM The optimal control problem is to minimize the objective function J ¼ 1 xT (N)Sx(N) þ 1 2 2 NÀ1 P k¼0 xT (k)Qx(k) þ uT (k)Ru(k) (5:50) subject to the constraint given by the state equations x(k þ 1) ¼ Ax(k) þ Bu(k) Using Lagrange multipliers.46).Closed-Loop System Analysis and Design 229 The state feedback control law is given by u(k) ¼ ÀK(k)x(k) (5:47) where K(k) is the controller feedback gain at time k.

l(k). we must have P(k) ¼ AT P(k þ 1)[I þ BRÀ1 BT P(k þ 1)]À1 A þ Q (5:63) (5:62) (5:61) (5:60) (5:59) (5:58) (5:57) (5:53) (5:54) (5:55) (5:56) The above equation can be simpliﬁed using the matrix inversion lemma which states that for any A. we need to differentiate H with respect to the three variables x(k). we have (A þ CD)À1 ¼ AÀ1 À AÀ1 C(I þ DAÀ1 C)À1 DAÀ1 (5:64) .57 into Equation 5.62 must be satisﬁed for all x(k).60 is now used to solve for x(k þ 1). That is x(k þ 1) ¼ [I þ BRÀ1 BT P(k þ 1)]À1 Ax(k) We now substitute Equation 5. we have x(k þ 1) ¼ Ax(k) À BRÀ1 BT P(k þ 1)x(k þ 1) Equation 5.51. C. we have u(k) ¼ ÀRÀ1 BT l(k þ 1) ¼ ÀRÀ1 BT P(k þ 1)x(k þ 1) By substituting Equation 5.53 results in Qx(k) þ AT P(k þ 1)x(k þ 1) À P(k)x(k) ¼ 0 From Equation 5.61 into Equation 5. and x(N) and set the results equal to zero. qH ¼ Qx(k) þ AT l(k þ 1) À l(k) ¼ 0 qx(k) qH ¼ Ru(k) þ BT l(k þ 1) ¼ 0 qu(k) qH ¼ Ax(k À 1) þ Bu(k À 1) À x(k) ¼ 0 ql(k) qH ¼ Sx(N) À l(N) ¼ 0 qx(N) The above equations can be solved by assuming that l(k) ¼ P(k)x(k) Substituting Equation 5. u(k). Therefore.54.58 to obtain Qx(k) þ AT P(k þ 1)[I þ BRÀ1 BT P(k þ 1)]À1 Ax(k) À P(k)x(k) ¼ 0 Equation 5.59 into Equation 5.230 Control of Color Imaging Systems: Analysis and Design To minimize the functional H. and D matrices of appropriate dimensions.

Closed-Loop System Analysis and Design 231 Using A ¼ I. the optimal control signal u(k) is given by u(k) ¼ ÀRÀ1 BT P(k þ 1)x(k þ 1) ¼ ÀRÀ1 BT P(k þ 1)[I þ BRÀ1 BT P(k þ 1)]À1 Ax(k) Therefore u(k) ¼ ÀK(k)x(k) where the time varying gain matrix K(k) is given by K(k) ¼ RÀ1 BT P(k þ 1)[I þ BRÀ1 BT P(k þ 1)]À1 A Example 5. we have P(k) ¼ AT P(k þ 1)A À AT P(k þ 1)BRÀ1 (I þ BT P(k þ 1)BRÀ1 )BT P(k þ 1)A þ Q (5:66) From Equation 5.66 with the boundary condition (Equation 5.65 into Equation 5. C ¼ BRÀ1 and D ¼ BT P(k þ 1) in matrix inversion lemma. Once P(k) is obtained. we have [I þ BRÀ1 BT P(k þ 1)]À1 ¼ I À BRÀ1 (I þ BT P(k þ 1)BRÀ1 )BT P(k þ 1) Substituting Equation 5.56 we have l(N) ¼ Sx(N) ¼ P(N)x(N) Therefore P(N) ¼ S (5:68) (5:67) (5:65) Equation 5.63. Find the optimal control feedback law to minimize the performance index J ¼ 2x2 (8) þ 1 2 7 P k¼0 (5:69) (5:70) (5:71) x2 (k) þ 2u2 (k) .68) is known as Riccati equation which can be solved backward in time starting from k ¼ N to k ¼ 0.6 Consider the control system given by x(k þ 1) ¼ 0:5x(k) þ u(k) with initial state x(0) ¼ 1.

232 Control of Color Imaging Systems: Analysis and Design SOLUTION In this problem A ¼ 0:5. Q ¼ 1.71 K(k) ¼ 0:25P(k þ 1)[1 þ 0:5P(k þ 1)]À1 ¼ Hence we have K(0) ¼ K(1) ¼ 0:25P(1) 0:25 Â 1:1861 ¼ ¼ 0:1861 1 þ 0:5P(1) 1 þ 0:5 Â 1:1861 0:25P(2) 0:25 Â 1:1861 ¼ ¼ 0:1861 1 þ 0:5P(2) 1 þ 0:5 Â 1:1861 0:25P(k þ 1) 1 þ 0:5P(k þ 1) .72 backward in time from k ¼ 8 to k ¼ 0 results in P(8) ¼ 4 P(7) ¼ P(6) ¼ P(5) ¼ P(4) ¼ P(3) ¼ P(2) ¼ P(1) ¼ P(0) ¼ 0:25P(8) 0:25 Â 4 þ1¼ þ 1 ¼ 1:3333 1 þ 0:5P(8) 1 þ 0:5 Â 4 0:25P(7) 0:25 Â 1:3333 þ1¼ þ 1 ¼ 1:200 1 þ 0:5P(7) 1 þ 0:5 Â 1:3333 0:25P(6) 0:25 Â 1:20 þ1¼ þ 1 ¼ 1:1875 1 þ 0:5P(6) 1 þ 0:5 Â 1:20 0:25P(5) 0:25 Â 1:1875 þ1¼ þ 1 ¼ 1:1863 1 þ 0:5P(5) 1 þ 0:5 Â 1:1875 0:25P(4) 0:25 Â 1:1863 þ1¼ þ 1 ¼ 1:1862 1 þ 0:5P(4) 1 þ 0:5 Â 1:1863 0:25P(3) 0:25 Â 1:1862 þ1¼ þ 1 ¼ 1:1861 1 þ 0:5P(3) 1 þ 0:5 Â 1:1862 0:25P(2) 0:25 Â 1:1861 þ1¼ þ 1 ¼ 1:1861 1 þ 0:5P(2) 1 þ 0:5 Â 1:1861 0:25P(1) 0:25 Â 1:1861 þ1¼ þ 1 ¼ 1:1861 1 þ 0:5P(1) 1 þ 0:5 Â 1:1861 The feedback control gain K(k) is computed using Equation 5. R ¼ 2. and N ¼ 8. S ¼ 4. The Riccati equation is P(k) ¼ 0:25P(k þ 1) À 0:125P(k þ 1)(1 þ 0:5P(k þ 1))À1 P(k þ 1) þ 1 This can be written as P(k) ¼ 0:25P(k þ 1) À 0:125P2 (k þ 1) 0:25P(k þ 1) þ1¼ þ1 1 þ 0:5P(k þ 1) 1 þ 0:5P(k þ 1) (5:72) with boundary condition P(8) ¼ S ¼ 4 Solving Equation 5. B ¼ 1.

the performance index is given by J¼1 2 1 P k¼0 xT (k)Qx(k) þ uT (k)Ru(k) (5:73) Under this scenario. From Equation 5.7 Consider the control system given by x(k þ 1) ¼ 0:5x(k) þ u(k) with initial state x(0) ¼ 1. the gain matrix K approaches a constant and is obtained by solving the steady-state Riccati equation.66.3.71 K ¼ RÀ1 BT P(I þ BRÀ1 BT P)À1 A Example 5.3 STEADY-STATE ALGEBRAIC RICCATI EQUATION When the time horizon N ! 1. the feedback gain matrix K is computed using Equation 5. Find the optimal control feedback law to minimize the performance index J ¼ þ1 2 1 P k¼0 (5:75) x2 (k) þ 2u2 (k) .Closed-Loop System Analysis and Design K(2) ¼ K(3) ¼ K(4) ¼ K(5) ¼ K(6) ¼ K(7) ¼ 0:25P(3) 0:25 Â 1:1862 ¼ ¼ 0:1862 1 þ 0:5P(3) 1 þ 0:5 Â 1:1862 0:25P(4) 0:25 Â 1:1863 ¼ ¼ 0:1863 1 þ 0:5P(4) 1 þ 0:5 Â 1:1863 0:25P(5) 0:25 Â 1:1875 ¼ ¼ 0:1875 1 þ 0:5P(5) 1 þ 0:5 Â 1:1875 0:25P(6) 0:25 Â 1:2 ¼ ¼ 0:2 1 þ 0:5P(6) 1 þ 0:5 Â 1:2 0:25P(7) 0:25 Â 1:3333 ¼ ¼ 0:3333 1 þ 0:5P(7) 1 þ 0:5 Â 1:3333 0:25P(8) 0:25 Â 4 ¼ ¼ 0:3333 1 þ 0:5P(8) 1 þ 0:5 Â 4 233 5. we have P ¼ AT PA À AT PBRÀ1 (I þ BT PBRÀ1 )BT PA þ Q (5:74) Once the positive deﬁnite matrix P is obtained.

5. Consider a MIMO system given as x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) where A 2 RNÂN B 2 RNÂM u 2 RM y 2 RP C 2 RPÂN The block diagram of the system is shown in Figure 5. 8. A state estimator is used to estimate the states of a dynamic system based on the output and the control signals [6. Hence.4. every state of a dynamic system is estimated. (5:76) (5:77) .4 STATE ESTIMATORS (OBSERVERS) DESIGN 5. we show the beneﬁts of the LQR design technique in designing high-performance color reproduction systems. Q ¼ 1. B ¼ 1. the order of the observer is the same as the order of the original system. In reality we may not have access to all the states.7]. The algebraic Riccati equation is P ¼ 0:25P À 0:125P(1 þ 0:5P)À1 P þ 1 ¼ or P2 þ 0:5P À 2 ¼ 0 The above algebraic quadratic equation has two solutions P ¼ 1:1861 and P ¼ À1:6861 0:25P þ1 1 þ 0:5P Since matrix P has to be positive deﬁnite.5.234 Control of Color Imaging Systems: Analysis and Design SOLUTION In this problem A ¼ 0:5. the acceptable solution is the positive solution.2 FULL-ORDER OBSERVER DESIGN In a full-order observer. and 10.1 INTRODUCTION In state feedback design using pole placement or optimal control using LQR. the assumption is that all the states are available for feedback. and R ¼ 2. 5.4. therefore the constant feedback gain is K¼ 0:25P 0:25 Â 1:1861 ¼ ¼ 0:1861 1 þ 0:5P 1 þ 0:5 Â 0:1861 In Chapters 7.

and eigenvalues are invariant under transpose operation. eigenvalues inside unit circle in complex z-plane).. Comparing the pole placement with this problem it is obvious that we can use any pole-placement algorithm with the following simple substitutions: A ! AT B ! CT K ! KT For example the Ackermann’s formula becomes n oT Â ÃÀ1 (NÀ1) 2 K ¼ ½ 0 0 Á Á Á 0 1 CT AT C T AT CT Á Á Á (A)T CT P(AT ) (5:82) (5:83) .e. Hence we need to choose gain matrix K by assigning eigenvalues to matrix A À KC.Closed-Loop System Analysis and Design x(k + 1) u(k) B + + 235 Unit delay x(k) C y(k) A FIGURE 5. The state observer is given by ^(k þ 1) ¼ A^(k) þ Bu(k) þ K ½y(k) À C^(k) x x x where ^ 2 RNÂN is an estimate of the state of the system x(k) x K is a gain matrix controlling the dynamic behavior of the observer Let the estimation error e(k) be e(k) ¼ x(k) À ^(k) x Then e(k þ 1) ¼ x(k þ 1) À ^(k þ 1) ¼ Ax(k) þ Bu(k) À A^(k) À Bu(k) À K[y(k) À C^(k)] x x x ¼ A(x(k) À ^(k)) À K[Cx(k) À C^(k)] ¼ (A À KC)(x(k) À ^(k)) x x x ¼ (A À KC)e(k) Equation 5.80 has a solution that is given by e(k) ¼ (A À KC)k e(0) (5:81) (5:80) (5:79) (5:78) Therefore if A À KC is a stable matrix (i.5 Block diagram of the system. Since A À KC ¼ (AT À C T K T )T . the problem is same as assigning eigenvalues to AT À C T K T . the error signal e(k) will approach zero as k ! 1. This is similar to the poleplacement algorithm discussed before. except that in pole placement we assign eigenvalues to A À BK.

and then K ¼ GT . 5 405 . 2 (5:85) x(k + 1) u(k) B + + Unit x(k) delay C y(k) A ˆ x (k) + + + Unit delay ˆ x (k) A ˆ (k) y C – + K FIGURE 5. Note that the necessary and sufﬁcient condition of the existence of a full-state observer is that observability matrix must have full rank. BT . that is 3 C 6 CA 7 6 7 6 2 7 rank6 CA 7 ¼ N 6 7 6 . 5 . 2 CANÀ1 1 (5:84) The MATLAB pole-placement command can also be used to compute the observer gain matrix as G ¼ place(AT . K ¼ P(A)6 7 6.6. . 7 6 7 4 . The block diagram of the full-order observer is shown in Figure 5.236 Control of Color Imaging Systems: Analysis and Design Equation 5.83 can be modiﬁed as 3À1 2 3 C 0 6 CA 7 6 0 7 6 7 6 7 6 7 6 7 6 CA2 7 6 .6 Block diagram of full-state observer.7 6 . 7 4 . P). CANÀ1 In other words the system must be full-state observable. 7 .

Place the poles of the observer at l1 ¼ l2 ¼ 0:25 SOLUTION The characteristic polynomial of the observer is P(l) ¼ (l À l1 )(l À l2 ) ¼ (l À 0:25)(l À 0:25) ¼ l2 À 0:5l þ 0:0625 The observability matrix is ! 1 C ¼ 0 CA 0 1 ! Since the observability matrix is full rank. 7 6 7 . the system is completely state observable. The observer gain vector is given by 3À1 2 3 C 0 6 CA 7 6 0 7 7 6 7 6 ! ! 7 6 7 6 C À1 0 6 CA2 7 6 . 7 ¼ [A2 À 0:5A þ 0:0625I] .Closed-Loop System Analysis and Design 237 Example 5. K ¼ P(A)6 7 6.7 1 CA 6 .8 Consider the dynamic system given by x(k þ 1) ¼ 0 À0:72 ! ! 1 0 x(k) þ u(k) 1:7 1 y(k) ¼ ½ 1 0 x(k) Design a full-state observer for the system. 5 405 1 CANÀ1 ! ! ! ! À0:6575 À0:33 1 0 0 À0:33 ¼ ¼ 0:2376 À0:7136 0 1 1 À0:7136 2 The state observer is given by ^(k þ 1) ¼ A^(k) þ Bu(k) þ K[y(k) À C^(k)] x x x ! ! ! 0 1 0 À0:33 ^(k þ 1) ¼ ^(k) þ x x u(k) þ (y(k) À ½ 1 À0:72 1:7 1 À0:7136 or ^(k þ 1) ¼ x ! ! ! 0 À0:33 1 0:33 ^(k) þ u(k) þ y(k) x 1 À0:7136 À0:0064 0:17 0 ^(k)) x . 4 .

A22 . A21 . Assume that the states of a dynamic system can be portioned into two parts. That is # " # " Measured x1 (k) (5:86) x(k) ¼ ÁÁÁ ¼ ÁÁÁ x2 (k) Not measured The state and output equations are portioned accordingly as x1 (k þ 1) ¼ A11 x1 (k) þ A12 x2 (k) þ B1 u(k) x2 (k þ 1) ¼ A21 x1 (k) þ A22 x2 (k) þ B2 u(k) y(k) ¼ Cx1 (k) where A11 . then its estimate is ^1 (k) ¼ x1 (k) ¼ C À1 y(k) x The second set of states x2 (k) are estimated as ^2 (k) ¼ Ly(k) þ z(k) x where z(k þ 1) ¼ Fz(k) þ Gy(k) þ Hu(k) The estimation error e(k) is deﬁned as e(k) ¼ x2 (k) À ^2 (k) x Dynamics of the estimation error e(k) is governed by e(k þ 1) ¼ x2 (k þ 1) À ^2 (k þ 1) ¼ A21 x1 (k) þ A22 x2 (k) þ B2 u(k) À Ly(k þ 1) À z(k þ 1) x ¼ A21 x1 (k) þ A22 x2 (k) þ B2 u(k) À LCx1 (k þ 1) À Fz(k) À Gy(k) À Hu(k) ¼ A21 x1 (k) þ A22 x2 (k) þ B2 u(k) À LC(A11 x1 (k) þ A12 x2 (k) þ B1 u(k)) À Fz(k) À GCx1 (k) À Hu(k) ¼ (A21 À LCA11 À GC)x1 (k) þ (A22 À LCA12 )x2 (k) þ (B2 À LCB1 À H)u(k) À Fz(k) ¼ (A21 À LCA11 À GC)x1 (k) þ (A22 À LCA12 )x2 (k) þ (B2 À LCB1 À H)u(k) (5:92) À F(^2 (k) À Ly(k)) x (5:87) (5:88) (5:89) (5:90) (5:91) . Consider the case when there are many states where most of them are measured except for a few. and C are matrices of appropriate size C is a nonsingular square matrix Since x1 (k) is measured. there is no need to estimate the available states.4.238 Control of Color Imaging Systems: Analysis and Design 5. A12 .3 REDUCED-ORDER OBSERVER DESIGN A full-order observer estimates all the states of the system regardless of whether they are measured or not. In this case. B2 . when partial measurement of some states is available. B1 . those that are measured directly and those that are not measured directly. A reduced-order observer is used to estimate the states of a dynamic system.

and H.Closed-Loop System Analysis and Design 239 Equation 5. The following simple example illustrates the design of a reduced-order observer: Example 5. we have F ¼ A22 À LCA12 H ¼ B2 À LCB1 G ¼ [A21 À LCA11 ]CÀ1 þ FL with these choices for F.92 can be simpliﬁed further as e(k þ 1) ¼ (A21 À LCA11 À GC)x1 (k) þ (A22 À LCA12 )x2 (k) x þ (B2 À LCB1 À H)u(k) À F^2 (k) þ FLCx1 (k) ¼ (A21 À LCA11 À GC þ FLC)x1 (k) þ (A22 À LCA12 )x2 (k) þ (B2 À LCB1 À H)u(k) À F(x2 (k) À e(k)) or e(k þ 1) ¼ Fe(k) þ (A21 À LCA11 À GC þ FLC)x1 (k) þ (A22 À LCA12 À F)x2 (k) þ (B2 À LCB1 À H)u(k) For the error to be independent of x1 (k). This is similar to the full-observer design with A22 À LCA12 playing the role of A À LC. x2 (k).9 Consider the dynamic system given by x(k þ 1) ¼ 0 À0:72 ! ! 0 1 u(k) x(k) þ 1 1:7 y(k) ¼ ½ 1 0 x(k) . and u(k). we must have A21 À LCA11 À GC þ FLC ¼ 0 A22 À LCA12 À F ¼ 0 B2 À LCB1 À H ¼ 0 Using the above three equations. The gain L matrix is designed by assigning eigenvalues to the matrix A22 À LCA12 . Equation 5. G.94 is reduced to e(k þ 1) ¼ Fe(k) ¼ (A22 À LCA12 )e(k) (5:99) (5:96) (5:97) (5:98) (5:94) (5:93) (5:95) The error signal will approach zero as k ! 1 if the eigenvalues of matrix A22 À LCA12 are inside unit circle in the complex plain.

The two processes of designing the state estimator and state feedback through pole placement are decoupled and are done separately. and B2 ¼ 1. A11 ¼ 0.240 Control of Color Imaging Systems: Analysis and Design Design a reduced-order observer to estimate the second state (x2 ) of the system. we have C ¼ 1. we have 1:7 À L ¼ 0:1 Solving for L yields L ¼ 1:6 and F ¼ 1:7 À L ¼ 0:1 H¼1 G ¼ À0:72 þ FL ¼ À0:72 þ 0:16 ¼ À0:56 Therefore. Place the pole of the reduced-order observer at l ¼ 0:1 SOLUTION Since y(k) ¼ x1 (k). B1 ¼ 0.2 COMBINED CONTROLLER AND OBSERVER The state feedback control law u(k) ¼ ÀKf x(k) is replaced by u(k) ¼ ÀKf ^(k) x (5:100) . a state feedback can be designed using the estimated states. A12 ¼ 1. Then F ¼ A22 À LCA12 ¼ 1:7 À L H ¼ B2 À LCB1 ¼ 1 G ¼ À0:72 þ FL To place the pole of the reduced-order observer at l ¼ 0:1. A22 ¼ 1:7.5 COMBINED STATE ESTIMATION AND CONTROL 5. the reduced-order observer is given as ^1 (k) ¼ x1 (k) ¼ y(k) x ^2 (k) ¼ 1:6y(k) þ z(k) x z(k þ 1) ¼ 0:1z(k) À 0:56y(k) þ u(k) 5.5. 5. A21 ¼ À0:72.1 INTRODUCTION Once the states of a dynamic system are estimated.5.

The block diagram of the overall system is shown in Figure 5.Closed-Loop System Analysis and Design 241 where ^(k) is the estimate of x(k) at time k x Kf is the state feedback gain matrix obtained through pole placement The combined controller and observer equations are given by x(k þ 1) ¼ Ax(k) þ Bu(k) ¼ Ax(k) À BKf ^(k) x ^(k þ 1) ¼ A^(k) þ Bu(k) þ Ke (y(k) À C^(k)) ¼ A^(k) À BKf ^(k) þ Ke (y(k) À C^(k)) x x x x x x (5:101) or ! x(k þ 1) A ¼ ^(k þ 1) Ke C x y(k) ¼ Cx(k) ÀBKf A À BKf À Ke C ! x(k) ^(k) x ! (5:102) where Ke is the observer gain matrix. . The estimation error is controlled by the eigenvalues of A À Ke C.7. If the (A. C) pair is observable.7 Combined state estimator and state feedback. the eigenvalues can be assigned arbitrarily similar to the pole x(k + 1) u(k) B + + Unit x(k) delay y(k) C ˆ x (k) A + + + ˆ (k) y Unit delay ˆ x (k) A C – + Ke –Kf FIGURE 5.

b1 ¼ À1:7 and b2 ¼ 0:72. a1 ¼ À0:5 and a2 ¼ 0:06. Example 5.242 Control of Color Imaging Systems: Analysis and Design placement.10 Consider the dynamic system given by x(k þ 1) ¼ 0 À0:72 ! ! 0 1 u(k) x(k) þ 1 1:7 y(k) ¼ ½ 1 0 x(k) Design a combined state feedback with a state observer for the system. the estimation error will approach zero faster than the dynamics of the system. The transformation T is given by T ¼ QW ¼ ½ B AB b1 1 ! ! À1:7 0 1 1 ¼ 1 1 1:7 0 ! 1 1 ¼ 0 0 0 1 ! The feedback gain matrix Kf is given as Kf ¼ ½ a2 À b2 a1 À b1 T À1 ¼ ½ À0:12 1 À1 0 0 1 !À1 ¼ ½ À0:66 1:2 We now design a full-state observer for the system. The characteristic polynomial of the desired closed-loop system is Pc (l) ¼ (l À 0:3)(l À 0:2) ¼ l2 À 0:5l þ 0:06 Hence. Place the eigenvalues of the closed-loop state feedback system at l1 ¼ 0:3 and l2 ¼ 0:2. Place the observer poles at l1 ¼ 0:1 and l2 ¼ 0:1. These eigenvalues should be located inside the unit circle closer to the origin than the eigenvalues of A À BKf . SOLUTION The characteristic polynomial of the open-loop system is P(l) ¼ jlI À Aj ¼ l2 À 1:7l þ 0:72 Hence. The characteristic polynomial of the observer is P(l) ¼ (l À l1 )(l À l2 ) ¼ (l À 0:1)(l À 0:1) ¼ l2 À 0:2l þ 0:01 The observability matrix is ! 1 C ¼ 0 CA 0 1 ! . With this eigenvalue assignment.

8 and 5. 7 6 7 4 .5 ˆ State x1(k) 1 0.Closed-Loop System Analysis and Design 243 Since the observability matrix is full rank. 3 State x1(k) 2.9. 7 ¼ A2 À 0:2A þ 0:01I 7 6. the system is completely state observable. 1 CANÀ1 ! ! ! ! À0:71 1:5 1 0 0 1:5 ¼ ¼ 1:08 1:84 0 1 1 1:84 2 C The dynamics of the state feedback and observer are given by ! A x(k þ 1) ¼ ^(k þ 1) x Ke C or 3 2 x1 (k þ 1) 0 6 x2 (k þ 1) 7 6 À0:72 7 6 6 4 ^1 (k þ 1) 5 ¼ 4 1:5 x ^2 (k þ 1) 1:84 x 2 1 0 1:7 0:66 0 À1:5 0 À1:9 32 3 x1 (k) 0 À1:2 76 x2 (k) 7 76 7 1 54 ^1 (k) 5 x ^2 (k) 0:5 x 2 T are shown in ÀBKf A À BKf À Ke C ! x(k) ^(k) x ! The states of the system with initial conditions of x(0) ¼ ½ 1 Figures 5. The observer gain vector Ke is given by 3À1 2 3 0 6 CA 7 6 0 7 7 6 7 6 ! ! 6 Â Ã C À1 0 2 7 6 . 7 Ke ¼ P(A)6 CA 7 6 . .7 6 CA 1 6 .8 Combined state estimator and state feedback (ﬁrst state). 5 405 .5 0 0 2 4 6 k 8 10 12 14 FIGURE 5.5 2 x1(k) 1.

we have (A þ CD)À1 ¼ AÀ1 À AÀ1 C(I þ DAÀ1 C)À1 DAÀ1 . Show that this system cannot be stabilized by the state feedback control u(k) ¼ ÀKx(k) for any choice of matrix K. x2 (k) ¼ y(k þ 1). l2 ¼ À0:4 þ j0:3.3 Prove matrix inversion lemma which states that for any A. obtain the gain matrix K so that the closed-loop poles are at l1 ¼ 0:1.9 Combined state estimator and state feedback (second state).5 ˆ State x2(k) 0 0 2 4 6 k 8 10 12 14 FIGURE 5. l3 ¼ À0:4 À j0:3 5. By use of state feedback control u ¼ ÀKx. 5. and x3 (k) ¼ y(k þ 2). and D matrices of appropriate dimensions.5 2 x2(k) 1. determine the uncontrollable mode. b. Derive state equation for this system.5 1 0. b. C. Deﬁne the state variables as x1 (k) ¼ y(k). Is it controllable? If not. PROBLEMS Y(z) 10z 5.244 3 Control of Color Imaging Systems: Analysis and Design State x2(k) 2.2 Consider the system deﬁned by ! ! ! ! À1 1 x1 (k) 1 x1 (k þ 1) ¼ þ u(k) 0 2 x2 (k) 0 x2 (k þ 1) a.1 A dynamic system has a plant: H(z) ¼ U(z) ¼ (zþ1)(zþ0:4)(zþ0:7) a.

6 Consider the system deﬁned by x1 (k þ 1) x2 (k þ 1) ! ¼ 0 1 ! x1 (k) x2 (k) ! þ 1 1 ! u(k) 0:4 0:5 y(k) ¼ x1 (k) Design a full-order state observer. The desired eigenvalues for the observer matrix are l1 ¼ 0:1 and 5.4 Consider the system deﬁned by x(k þ 1) ¼ 0:9x(k) þ u(k) Find the state feedback control law u(k) ¼ ÀK(k)x(k) minimizing the performance index J ¼ 2x2 (10) þ 1 2 5.Closed-Loop System Analysis and Design 245 5. À0:4 2 3 0 B ¼ 405 1 l2 ¼ 0:2 .5 Consider the system deﬁned by ! ! ! ! À0:6 1 x1 (k) 1 x1 (k þ 1) ¼ þ u(k) 0 0:5 x2 (k) 0 x2 (k þ 1) Find the state feedback control law u(k) ¼ ÀKx(k) minimizing the performance index J¼1 2 1 P k¼0 9 P k¼0 [x2 (k) þ 0:5u2 (k)] x2 (k) þ 2x2 (k) þ u2 (k) 1 2 5.7 Consider the system deﬁned by x(k þ 1) ¼ Ax(k) þ Bu(k) where 0 1 A¼4 0 0 0:1 0:25 2 3 0 1 5.

The desired eigenvalues for the reduced-order observer matrix are l1 ¼ l2 ¼ 0:1 5.10 Consider the system x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) where 0 A¼4 0 0:1 2 1 0 0:25 3 2 3 0 1 0 1 5.9 Consider the system deﬁned by x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) where 0 A¼4 0 0:1 2 3 2 3 1 0 0 0 1 5.8 Consider the system deﬁned by x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) where 0 A¼4 0 0:1 2 3 2 3 1 0 0 0 1 5. B ¼ 4 0 5. Choose the desired closed-loop poles for deadbeat control.246 Control of Color Imaging Systems: Analysis and Design Design a linear state feedback for this system. C ¼ 1 À0:4 1 0 0 1 0 ! . The desired eigenvalues for the observer matrix are l1 ¼ l2 ¼ 2l3 ¼ 0:2 5. B ¼ 4 0 5. 5. 0:25 À0:4 1 C ¼ ½1 0 0 Design a full-order state observer. 0:25 À0:4 1 C ¼ ½1 0 0 Design a reduced-order state observer. B ¼ 4 0 5.

Modern Control Theory. 1990. 7. Tewari. Naeini. 1985. A. REFERENCES 1. L. Ogata. Modern Control Engineering. K. which is the desired eigenvalue of the reduced-order observer. G. B. Englewood. NJ. Ogata. New York. McGrawHill. Englewood. Prentice Hall. Powell. NJ. 6. 1994. and E. 2001. Wiley. 4. . which is l ¼ 0. K. NJ. 5. New York. Brogan. Friedland. D. United Kingdom.Closed-Loop System Analysis and Design 247 Design a reduced-order state observer such that the response to the observer error is deadbeat. PrenticeHall. W. Prentice Hall. Control System Design: An Introduction to State-Space Methods. 2. NJ. Englewood. Upper Saddle River. Franklin. Prentice Hall. 3. Modern Control Design with MATLAB and Simulink. B. Digital Control Systems. Chichester. Feedback Control of Dynamic Systems. J. Discrete-Time Control Systems. 1995. 2002. Oxford University Press. 2005. Kuo.

.

for example. After this. we have two options. a numerical computation of inverse printer maps and techniques for downsampling the color space LUTs are also presented in some detail. K) Unfortunately. The ﬁrst option is to have a lookup table (LUT) of all possible combinations of CMYK. An extension to 4-D color space is straightforward. M. The mapping can be deﬁned by the following three multidimensional functions: L* ¼ f1 (C. Y. and inverse printer maps. There are two types of interpolation techniques: linear and nonlinear. Therefore. Y. which requires a LUT size of 2554 entries that is equivalent to 3 Â 2554 bytes ¼ 12. Multidimensional interpolation is also extensively used in displays. we present different methods of multidimensional interpolation.6856 bytes of CMYK – L*a*b* data. change each separation from 0 to 255 in steps of one. two-dimensional (2-D). Another approach would be to have a LUT of size smaller than the full size. 174 or lesser. that is. For example. we consider nonuniform LUTs and describe nonlinear techniques such as the Shepard and moving-matrix approaches. M. and 3-D color spaces.6 Interpolation of Multidimensional Functions 6. Y.1 INTRODUCTION Multidimensional functions play an important role in digital color imaging systems. As an example of application. This is not practical. scanners. linear interpolation techniques are used to interpolate the data. CMYK – L*a*b* data and use interpolation to ﬁnd other colors. K) a* ¼ f2 (C. K) b* ¼ f3 (C. where the color space is uniformly sampled (uniform LUT) and is reasonably linear. 249 (6:1) . exact closed-form expressions are not easily available for the above equations. then nonlinear interpolations are preferred. a digital color printer can be modeled as a forward map from fourdimensional (4-D) device-dependent CMYK color space to three-dimensional (3-D) device-independent L*a*b* color space. If the color space data is nonuniform and is not linear. that is. In this chapter. Generally speaking. We ﬁrst consider uniform LUTs and discuss linear interpolation in onedimensional (1-D). M.

1 LINEAR AND BILINEAR INTERPOLATIONS We ﬁrst consider the linear interpolation of a 1-D LUT. . Trilinear interpolation with application to digital printers is discussed in the subsequent section.250 Control of Color Imaging Systems: Analysis and Design 6. as shown in Figure 6. shown in Figure 6. xi x xiþ1. .2. xN y ¼ f(x) y1 y2 . TABLE 6. 6. that is.1 Uniform LUT of N Data Points x x1 x2 .1. yN . The interpolated value y is given by y¼ for xi x xiþ1.1 One-dimensional function f(x). The interpolated value is located on the straight line connecting the two nodes a and c. f (x) y1 yN yiþ1 À yi ð x À xi Þ þ yi xiþ1 À xi (6:2) yi + 1 y=? yi x1 b a xi x c xi+ 1 xN x FIGURE 6. Assume that a uniform LUT of N data points is available. cubic. This point is located between the two LUT nodes a and c.1.1. and other nonlinear techniques [1].2 INTERPOLATION OF UNIFORMLY SPACED LOOKUP TABLES Interpolation of uniformly spaced LUTs can be performed using linear. . Consider the 1-D function y ¼ f(x). where x is the independent variable and y is the dependent variable. The LUT is shown in Table 6. . We will ﬁrst consider linear (1-D) and bilinear (2-D) interpolation techniques. Suppose that we would like to interpolate point b on the curve.

p01. Use linear interpolation to interpolate the value of the function at x ¼ 5. y). the underlying function is a function of two variables.4. Let point a be within the grids b–e as shown in Figure 6. y) at points b.2 Plot of function f(x). y).5 10 11. c. Then the bilinear interpolated value of the function at point a is given by z ¼ f (x. SOLUTION Figure 6. and p10 are the values of the function f(x.4. p11.2 shows the plot of function y ¼ f(x). and e. Assume that we would like to interpolate the value of the function z at point a with coordinates (x. .4.2 Example of a Uniform LUT x y 0 0.98 4 1. TABLE 6. as shown in Figure 6. respectively. as shown in Figure 6.8 12 10 8 6 4 2 0 y 0 1 2 3 4 5 x 6 7 8 9 10 FIGURE 6.3 8 6.2.Interpolation of Multidimensional Functions 251 Example 6. z ¼ f(x. we have y¼ yiþ1 À yi 3:3 À 1:62 ðx À xi Þ þ yi ¼ (5:4 À 6) þ 3:3 ¼ 2:796 xiþ1 À xi 6À4 In the 2-D case. and the nodes in the LUT are uniformly spaced on grids.1 Consider the 1-D LUT given in Table 6. As seen from this ﬁgure. d.62 6 3.3.9 2 0. y) ¼ p00 þ t ðp01 À p00 Þ þ uðp10 À p00 Þ þ tuðp11 À p01 À p10 À p00 Þ (6:3) where u and t are the relative distances from the surrounding nodes p00.

and two outputs. p01 p11 1–t 1– u u t p00 p10 FIGURE 6. yi Þ À f ðxi .5 and x2 ¼ 0. y) ¼ f ðxi . x1 and x2. yi Þ À f ðxi . yi Þ xiþ1 À xi yiþ1 À yi (6:5) Example 6.3 and Figure 6. The LUT representing these two functions is given in Table 6.3 results in x À xi y À yi z ¼ f (x.3 Two-dimensional uniform grid. The relative distances u and t are given by u¼ x À xi . where y1 ¼ f1(x1. yiþ1 Þ À f ðxiþ1 . Use bilinear interpolation to ﬁnd the value of the two functions at point x ¼ [x1 x2].2 Consider a static system with two inputs. yi Þ þ ½ f ðxi . yiþ1 Þ À f ðxi .5. where x1 ¼ 1. yi Þ xiþ1 À xi yiþ1 À yi x À xi y À yi þ ½ f ðxiþ1 . x2).252 y yN c y b y2 y1 x1 x2 x a Control of Color Imaging Systems: Analysis and Design y d e yi+1 y yi c a b d e xN x xi x xi + 1 x FIGURE 6. xiþ1 À xi t¼ y À yi yiþ1 À yi (6:4) Substituting u and t in Equation 6.6.4 Relative distances. y1 and y2. yiþ1 Þ þ f ðxi . yi Þ þ ½ f ðxiþ1 . . x2) and y2 ¼ f2(x1.

5 Range and domain of the two functions y1(x1. c¼ b¼ 0 ! 2 . and x2 is between 0 and 1. and e ¼ . x2).Interpolation of Multidimensional Functions 253 TABLE 6. p11 ¼ . x2) and y2(x1. d¼ 0 1 1 ! ! ! 3 5 4 . and p10 ¼ 4 5 2 p00 . the four surrounding points ! y1 values are in x1 À x2 space and their corresponding y2 ! 1 . p01 ¼ ¼ 1 ! ! ! 2 2 1 . SOLUTION The values of u and t are u¼ 1:5 À 1 ¼ 0:5 and 2À1 t¼ 0:6 À 0 ¼ 0:6 1À0 since x1 is between 1 and 2.3 LUT for a Two-Input Two-Output System x1 0 0 0 1 1 1 2 2 2 x2 0 1 2 0 1 2 0 1 2 y1 0 1 2 2 3 4 4 5 6 y2 0 3 6 1 4 7 2 5 8 x2 y2 2 c 1 x2 b 0 x1 e x1 d y2 p01 p00 p11 p10 y1 y1 FIGURE 6.

As can be seen. z). and v be the relative distances from the plane and the axes. z y x FIGURE 6. the underlying function is a function of three variables.6. u. Let the surrounding eight points in the xyz plane be ½ n000 n001 n010 n011 n100 n101 n110 n111 and the corresponding points in the f plane be ½ p000 p001 p010 p011 p100 p101 p110 p111 . there are a total of 64 grid points uniformly spaced in the 3-D space. Let t.2.254 Therefore. z). y1 y2 ! Control of Color Imaging Systems: Analysis and Design ¼ p00 þ tðp01 À p00 Þ þ uðp10 À p00 Þ þ tuðp11 À p01 À p10 À p00 Þ ! ! ! ! ! 2 3À2 4À2 5À4À3þ2 3:6 ¼ þ 0:6 þ 0:5 þ 0:6 Â 0:5 ¼ 1 4À1 2À1 5À2À4þ1 3:3 We next consider a trilinear interpolation of 3-D LUTs.6 Uniformly sampled lattice in 3-D space.7. y.2 TRILINEAR INTERPOLATION In the 3-D case. 6. f(x. as shown in Figure 6. An example of such a lattice for N ¼ M ¼ L ¼ 4 is shown in Figure 6. y. Now assume that we have a 3-D LUT of size N Â M Â L grid points uniformly spaced in the xyz 3-D space and would like to interpolate the grid point (x. as shown in Figure 6.7. . and the LUT is uniformly spaced in a 3-D lattice of N Â M Â L points.

as shown in Figure 6. The interpolated value of f at location (x. v.7: p0 ¼ p00 þ vðp10 À p00 Þ p1 ¼ p01 þ vðp11 À p01 Þ (6:7) (6:6) (3) Use linear interpolation to interpolate the corresponding values of the function f at nodes [x y z]. z) is computed in three steps as follows: (1) Use bilinear interpolation to interpolate the corresponding values of the function f at nodes ½ n00 n01 n10 n11 . as shown by the dotted plane in Figure 6. z) pijk n10 n110 n0 n000 n00 n100 FIGURE 6.Interpolation of Multidimensional Functions xyz plane 255 n011 t n01 n001 v n010 u n1 n11 n111 n101 f (x.7: p00 ¼ p000 þ t ðp100 À p000 Þ p01 ¼ p001 þ t ðp101 À p001 Þ p10 ¼ p010 þ t ðp110 À p010 Þ p11 ¼ p011 þ t ðp111 À p011 Þ (2) Use linear interpolation to interpolate the corresponding values of the function f at nodes [n0 n1]. and yiþ1 À yi u¼ z À zi ziþ1 À zi (6:9) (6:8) . and u are given by t¼ x À xi . y. y. xiþ1 À xi v¼ y À yi . as shown in Figure 6.7 Interpolation cube and its eight vertices.7: p ¼ p0 þ uðp1 À p0 Þ The relative distances t.

The LUT representing the function is given in Table 6. TABLE 6. TABLE 6. The point p ¼ [x y z]T ¼ [1:5 0:6 1:85]T is inside the cube with vertices ½ n000 n001 n010 n011 n100 n101 n110 n111 . y.6.5.256 Control of Color Imaging Systems: Analysis and Design Example 6.85.8 Eight vertices of the cube surrounding the point p. and z ¼ 1.3 n000 x y z f 1 0 1 3 n001 1 0 2 2 n010 1 1 1 2 n011 1 1 2 4 n100 2 0 1 2 n101 2 0 2 5 n110 2 1 1 1 n111 2 1 2 3 .5.5 LUT for Example 6. and z. x. The coordinates of these nodes and the corresponding values of the function f are given in Table 6. as shown in Figure 6. Use trilinear interpolation to ﬁnd the value of the function at points x ¼ 1.4.8.4 Example of a Uniformly Sampled 3-D LUT x y z f 0 0 0 2 1 0 0 3 2 0 0 4 0 1 0 1 1 1 0 5 2 1 0 7 0 2 0 8 1 2 0 3 2 2 0 0 0 0 1 4 1 0 1 3 2 0 1 2 0 1 1 3 1 1 1 2 2 1 1 1 0 2 1 3 1 2 1 7 2 2 1 8 0 0 2 9 1 0 2 2 2 0 2 5 0 1 2 6 1 1 2 4 2 1 2 3 0 2 2 8 1 2 2 6 2 2 2 3 n011 n11 n1 n001 n01 p n101 n111 n010 n10 n000 n0 n00 n100 n110 FIGURE 6.3 Consider a system with three inputs. and one output. f. y ¼ 0.

p ¼ p000 þ px x À x0 y À y0 z À z0 þ py þ pz x1 À x0 y1 À y0 z1 À z0 (6:10) The expressions for px. then p is in tetrahedral 1 and px ¼ p100 À p000 . py ¼ p110 À p100 . the interpolated value of the function at point p is p ¼ p0 þ uðp1 À p0 Þ ¼ 1:9 þ 0:85(3:5 À 1:9) ¼ 3:26 6. Now assume that we have a 3-D LUT of size N Â M Â L grid points uniformly spaced in xyz 3-D space and would like to interpolate the grid point (x. and u are t¼ v¼ u¼ x À xi 1:5 À 1 ¼ ¼ 0:5 xiþ1 À xi 2À1 y À yi 0:6 À 0 ¼ ¼ 0:6 yiþ1 À yi 1À0 z À zi 1:85 À 1 ¼ ¼ 0:85 ziþ1 À zi 2À1 n01 n10 n11 are The interpolated values at nodes ½ n00 p00 p01 p10 p11 ¼ p000 þ tðp100 À p000 Þ ¼ 3 þ 0:5(2 À 3) ¼ 2:5 ¼ p001 þ tðp101 À p001 Þ ¼ 2 þ 0:5(5 À 2) ¼ 3:5 ¼ p010 þ tðp110 À p010 Þ ¼ 2 þ 0:5(1 À 2) ¼ 1:5 ¼ p011 þ tðp111 À p011 Þ ¼ 4 þ 0:5(3 À 4) ¼ 3:5 n1 are The interpolated values at nodes ½ n0 p0 ¼ p00 þ vðp10 À p00 Þ ¼ 2:5 þ 0:6(1:5 À 2:5) ¼ 1:9 p1 ¼ p01 þ vðp11 À p01 Þ ¼ 3:5 þ 0:6(3:5 À 3:5) ¼ 3:5 Finally.9. py. and pz ¼ p111 À p110 (6:11) .3 TETRAHEDRAL INTERPOLATION Tetrahedral interpolation is another approach for interpolating the regularly sampled LUTs [2]. The tetrahedral interpolation divides this cube into six tetrahedrals. y. The interpolated value is the weighted sum of the values of the function at the four vertices of the tetrahedral enclosing the desired point. as shown in Figure 6. as shown in Figure 6.2.8. z). and pz depend on the location of p with respect to the six tetrahedral and are given as follows: (1) If x À x0 > y À y0 > z À z0 .Interpolation of Multidimensional Functions 257 SOLUTION The parameters t. That is. Let the surrounding eight nodes in the xyz plane be ½ n000 n001 n010 n011 n100 n101 n110 n111 and the corresponding points in the f plane be ½ p000 p001 p010 p011 p100 p101 p110 p111 . v.

(2) If x À x0 > z À z0 > y À y0 .9 Dividing the cube into six tetrahedrals. then p is in tetrahedral 6 and px ¼ p111 À p011 . then p is in tetrahedral 5 and px ¼ p111 À p011 . py ¼ p111 À p101 . then p is in tetrahedral 4 and px ¼ p110 À p010 . py ¼ p010 À p000 .258 Control of Color Imaging Systems: Analysis and Design Tetrahedral 1 Tetrahedral 2 Tetrahedral 3 Tetrahedral 4 Tetrahedral 5 Tetrahedral 6 FIGURE 6. then p is in tetrahedral 2 and px ¼ p100 À p000 . and pz ¼ p001 À p000 (6:13) (4) If y À y0 > x À x0 > z À z0 . and pz ¼ p011 À p010 (6:15) (6) If z À z0 > y À y0 > x À x0 . py ¼ p011 À p001 . py ¼ p111 À p101 . py ¼ p010 À p000 . and pz ¼ p111 À p110 (6:14) (5) If y À y0 > z À z0 > x À x0 . then p is in tetrahedral 3 and px ¼ p101 À p001 . and pz ¼ p101 À p100 (6:12) (3) If z À z0 > x À x0 > y À y0 . and pz ¼ p001 À p000 (6:16) .

00 97.28 24.21 41.33 41.18 78.16 83.79 48.16 70.84 78.02 94. 85 2 3 0 7 6 ¼ 4 170 5.82 54.61 À26.50 a* À0.98 81.18 87.10 68. The CMY values at the eight vertices of the cube are 3 0 6 7 ¼ 4 85 5.80 53. 0 2 3 0 7 6 ¼ 4 170 5.49 49.4 Consider the transformation from device-dependent CMY to device-independent L*a*b* given by the LUT shown in Table 6. 0 2 3 0 6 7 ¼ 4 85 5.6. 0 3 2 85 7 6 ¼ 4 170 5.53 À8.69 À24.95 55.96 82.72 44.95 À11.90 68.00 À7.07 15.66 68. 0 2 3 85 6 7 ¼ 4 85 5.85 À14. 85 2 3 85 6 7 ¼ 4 85 5.64 À20.16 86.13 79.47 43.65 29.21 69. 85 2 2 n111 n000 n001 n010 n011 n100 n101 n110 and 3 85 7 6 ¼ 4 170 5 85 TABLE 6.25 9.03 À20.10.72 19.25 73.07 55. as shown in Figure 6.76 16.72 100. Using tetrahedral interpolation.59 b* 0.59 29.41 À25.54 77.42 (continued) .70 53.04 À4.49 80.00 35.88 À7.86 23.44 12.02 À13.6 Example of a Printer Forward Map (CMY to L*a*b*) C 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 85 85 85 85 85 M 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 0 0 0 0 85 Y 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 L* 100.Interpolation of Multidimensional Functions 259 Example 6. ﬁnd the L*a*b* values corresponding to the CMY ¼ [67 128 54] color patch.51 69.94 80.82 17.58 55.85 69.86 68.61 96. SOLUTION The desired CMY ¼ [x y z] ¼ [67 128 54] point is inside the tetrahedral.70 81.

28 35.03 45.54 44.96 53.53 25.92 54.81 2.64 36.18 À51.57 13.49 54.50 À24.40 57.12 30.31 À45.77 14.91 9.25 À57.30 1.20 34.74 À17.13 À24.40 À30.44 À11.20 53.48 37.60 31.60 66.56 54.02 32.35 À28.31 40.56 À16.02 .44 26.79 À61.89 À47.24 À18.45 45.98 À9.77 À55.51 À30.78 À34.10 11.67 À35.15 À16.88 5.19 À46.01 27.90 À43.53 45.99 À14.64 23.35 47.00 62.72 56.6 (continued) Example of a Printer Forward Map (CMY to L*a*b*) C 85 85 85 85 85 85 85 85 85 85 85 170 170 170 170 170 170 170 170 170 170 170 170 170 170 170 170 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 M 85 85 85 170 170 170 170 255 255 255 255 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 Y 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 L* 67.76 54.25 13.71 À10.79 54.42 À5.37 63.91 À3.96 À39.83 À0.95 À6.17 33.75 27.50 À55.55 b* 5.46 0.14 À22.26 À30.37 23.01 54.82 66.22 À57.72 À12.53 37.06 6.77 40.21 À2.39 45.40 39.98 26.68 23.42 13.37 61.72 À56.20 38.65 56.80 30.05 4.42 À1.73 67.76 44.99 28.92 11.65 1.92 49.82 À41.67 57.73 23.51 À27.12 24.80 32.47 70.14 57.65 54.66 À38.95 18.10 35.94 46.67 37.07 34.30 65.47 À44.04 a* 7.67 45.60 45.69 54.93 À29.12 42.10 44.11 64.260 Control of Color Imaging Systems: Analysis and Design TABLE 6.89 42.41 45.84 44.48 À7.

p101 ¼ 4 7:81 5. p010 ¼ 4 55:79 5.Interpolation of Multidimensional Functions nijk pijk n101 n111 261 n011 n001 80 70 60 M 50 40 30 20 10 0 0 n000 20 40 60 80 100 n100 n 010 Y C n110 150 FIGURE 6. À7:28 24:18 À11:76 16:33 3 3 3 3 2 2 2 2 68:50 67:73 56:67 57:40 7 7 7 7 6 6 6 6 p100 ¼ 4 15:59 5. p011 ¼ 4 48:72 5. The corresponding L*a*b* values are ½ p000 p001 p010 p011 p100 p101 p110 p111 . p110 ¼ 4 40:60 5. and pz vectors are 3 2 3 2 3 68:50 82:98 À14:48 px ¼ p100 À p000 ¼ 4 15:59 5 À 4 29:86 5 ¼ 4 À14:27 5 À25:42 À7:28 À18:14 2 3 2 3 2 3 57:40 67:73 À10:33 py ¼ p111 À p101 ¼ 4 31:75 5 À 4 7:81 5 ¼ 4 23:94 5 À0:64 5:02 À5:66 2 3 2 3 2 3 67:73 68:50 À0:77 pz ¼ p101 À p100 ¼ 4 7:81 5 À 4 15:59 5 ¼ 4 À7:78 5 5:02 À25:42 30:44 2 .10 Tetrahedral surrounding the target color. p001 ¼ 4 23:82 5. which are obtained directly from the LUT: 3 3 3 3 2 2 2 82:98 81:49 69:10 68:90 7 7 7 7 6 6 6 6 p000 ¼ 4 29:86 5. and p111 ¼ 4 31:75 5 À25:42 5:02 À28:83 À0:64 2 The px. py.

if one can optimally place the grid points to achieve good approximations of the multidimensional functions.2. then the color space is optimally utilized and the resulting mean-square error (MSE) due to interpolation will be small. at points pR and pL. therefore. a 3-D uniformly sampled L*a*b* to CMY LUT used to characterize a digital printer will have many colors that are outside the printer gamut. Figure 6. we consider 2-D functions and extend the results to 3-D.11 shows an SLI grid structure LUT of size 19. the approach is similar to linear interpolation.4 SEQUENTIAL LINEAR INTERPOLATION Uniformly sampled LUTs do not use the color space efﬁciently.262 Hence. x2) be a nonlinear function of the two variables x1 and x2.12. yR and yL. p ¼ p000 þ px Control of Color Imaging Systems: Analysis and Design x À x0 y À y0 z À z0 þ py þ pz x1 À x0 y1 À y0 z1 À z0 3 3 3 3 2 2 2 2 82:98 À14:48 À10:33 À0:77 7 67 À 0 6 7 128 À 85 6 7 54 À 0 6 7 6 7 7 7 6 6 6 ¼ 6 29:86 7 þ 5 85 À 0 Â 4 À14:27 5 þ 170 À 85 Â 4 23:94 5 þ 85 À 0 Â 4 À7:78 5 4 À7:28 À18:14 À5:66 30:44 Therefore. For example. 2 65:85 3 7 6 p ¼ 6 25:78 7 5 4 À5:10 6. They are given by yR ¼ wR f ðpR1 Þ þ ð1 À wR Þ f ðpR2 Þ and yL ¼ wL f ðpL1 Þ þ ð1 À wL Þ f ðpL2 Þ (6:18) (6:17) . The sequential linear interpolation (SLI) is an optimal approach for approximating multidimensional functions by selecting the location of the grid points in a sequential form suitable for sequential interpolation [3]. Therefore. Let y ¼ f(x) ¼ f(x1. Here we assume that the grid points have already been selected optimally according to the SLI algorithm and only discuss the SLI. respectively. One-dimensional linear interpolation is used to estimate the values of the function. For the 1-D functions. as shown in Figure 6. To interpolate a grid point p ¼ [x1 x2] that is not part of the LUT. we ﬁrst project this point to the two nearest grid lines to the right and left of p.

x2 pL2 p pL pL1 pR2 pR pR1 x1L x1R x1 FIGURE 6. where wR ¼ x2 À x2R1 x2 À x2L1 and wL ¼ . Once yR and yL are determined. .12 Interpolation of point p lying between grid line x1 ¼ x1L and x1 ¼ x1R.Interpolation of Multidimensional Functions x2 263 b24 b23 b11 b22 b21 a1 a2 a3 a4 a5 a6 x1 FIGURE 6. the x2R2 À x2R1 x2L2 À x2L1 value of the function at point p is obtained by y ¼ uyL þ (1 À u)yR where u¼ x1 À x1R x1L À x1R (6:20) (6:19) The above technique can be easily extended to 3-D and 4-D LUTs with SLI structures.11 Example of a 2-D sequential grid structure.

3.3 NONUNIFORMLY SPACED LOOKUP TABLES There are several algorithms available to interpolate nonuniformly spaced LUTs. However. None of these techniques provide accurate estimations of the values of the underlying functions at grid points that are not part of the LUT. indicating that the y interpolation formula given in Equation 6.1 SHEPARD INTERPOLATION The Shepard interpolation is an approach to interpolate multidimensional irregularly spaced data [4]. The function selected is continuously differentiable. Let the set be given by S ¼ f(x1 . p and m. that can be selected by the user. yN )g (6:22) where xi 2 RM and yi 2 RM. The Shepard algorithm is based on a section of a local function to approximate the underlying function at a given point using weighted regression. Then the interpolated value at x is given by 8 PN Àm > i¼1 yi di < PN > : i¼1 ^¼ y diÀm if d i 6= 0 if d j = 0 for all i for some j (6:23) yj where di is the distance (based on Lp norm) between vectors x and xi. To discuss the algorithm. that is. then ^ ! y ¼ P(x).6) that if the parameter m is chosen to be greater than 1. and the dynamic least-square interpolation algorithm. There are two parameters. Let x be a point in the input space with the corresponding point y in the output space. Among them are the Shepard interpolation algorithm. It can be shown (see Problem 6. as desired. 6.264 Control of Color Imaging Systems: Analysis and Design 6. where x À y. the moving-matrix technique. y1 ) (x2 . .23 is a continuous function. y2 ) Á Á Á (xN . consider a LUT that maps M-dimensional vector y to M-dimensional vector z through a nonlinear transformation. di ¼ kx À xi kp (6:24) Note that if x is approaching a data point. the interpolated function will be differentiable. they are the only techniques known to us that can be used for the interpolation of nonuniformly spaced LUTs that are not structured. ! P y ¼ P(z) (6:21) Assume that a set of N irregularly spaced data points are available. These techniques are discussed in the following sections.

66 0.0334 0.5 Consider the 1-D function y ¼ p(x) deﬁned by p(x) ¼ x2 þ x3 sin (4px) þ 2 24x 0 x 1 We ﬁrst nonuniformly sample this function at N ¼ 10 points and create the LUT given in Table 6.4 0. (a) Use the Shepard interpolation to estimate the value of the function at x ¼ 0.9 1 FIGURE 6.27 0.0079 0. Use m ¼ 2 and p ¼ 2.419 0.4 0.80 0.3 0.41 0.0266 0.2 0.55 0.8 0.7 0.13 Shepard interpolation of a 1-D function.7. TABLE 6.2 1 0. (b) Repeat part a for 100 samples of x that are uniformly spaced between 0 and 1.8 0.1 0.6 f (x) 0.545 0.07 0.89 0.2 0 –0.2 Original f (x) Irregularly spaced samples Shepard interpolation 0 0.5 x 0.7 Uniformly Sampled LUT at N ¼ 10 Points x y 0 0.6 0.523 0. and the interpolated function is shown in Figure 6. the irregularly sampled points.Interpolation of Multidimensional Functions 265 Example 6. Plot the results and compare them with the exact values of y.279 0. .5.702 1 1 1.461 0.13. SOLUTION PN P10 À2 P10 P10 y d Àm yd y jxÀxi jÀ2 y j0:5Àxi jÀ2 i¼1 i i i¼1 i i i¼1 i i¼1 i ^ (a) y ¼ PN Àm ¼ P10 À2 ¼ P10 ¼ P10 ¼ 0:2292 À2 À2 i¼1 di i¼1 di i¼1 jxÀxi j i¼1 j0:5Àxi j (b) The plot of the original function.34 À0.

8.55 b* 0.38 37.04 75.5 127.93 40. and b* ¼ 40.87 À61.5 255 0 127.02 44.89 À35.27 59.12 53.30 100.5 255 0 127.5 127.69 55.13 À9.38 À8.49 94.266 Control of Color Imaging Systems: Analysis and Design Example 6.48 24.5 255 255 255 0 0 0 127.06 8.68 À31.56 54.09 19.93 30.53 86.5 255 0 127.91 34.5 127.09 70. Use m ¼ 2 and p ¼ 2.37 55.5 127.04 À5.67 À26.15 À22.5 127.02 .35 40.12 34.82 3.62 42.72 70.5 255 255 255 0 0 0 127.04 74.5 255 0 127.5 255 0 127.18 75.86 49.5 255 0 127.69 À38. Use m ¼ 2 and p ¼ 2.5 127.5 127.91 À57.5 255 L* 100 94.96 75.5 127.5 127.37 54.00 50.52 À15.5 255 0 127.04 a* À0.54 39. a* ¼ 80.03 À29.46 24.59 43.98 26.09 40.70 54.6 Consider a CMY printer with the 3 Â 3 Â 3 forward printer map LUT (CMY ! L* a* bÁ*) given in Table 6.43 5.27 À40.49 55.5 127. Use the Shepard interpolation algorithm to ﬁnd (a) The CMY corresponding to L* ¼ 50.5 255 255 255 255 255 255 255 255 255 M 0 0 0 127.61 41.02 À21.92 42.23 1.17 28.59 À4.5 127.00 À17. (b) The CMY corresponding to L* ¼ 50.82 55.5 127. and b* ¼ À14.41 74.55 À55. SOLUTION PN P27 P27 Àm À2 CMYi k[50 30 40] À L*a*bi *kÀ2 i¼1 yi di i¼1 yi di ^ (a) y ¼ PN Àm ¼ P27 À2 ¼ i¼1 27 P À2 i¼1 di i¼1 k[50 30 40] À L*a*bi *k i¼1 di ¼ [98:27 161:53 176:64] TABLE 6.77 À57.70 16.5 255 0 127.5 255 255 255 Y 0 127. a* ¼ 30.8 3 Â 3 Forward Printer MAP C 0 0 0 0 0 0 0 0 0 127.00 À8.27 59.71 50.72 À2.03 23.18 À54.65 À37.79 29.27 71.96 À14.5 127.70 34.25 20.16 73.88 17.41 56.5 127.

[C M Y] ¼ [27:35 240:57 26:81] 240:57 26:81] 6. then the number of terms in the augmented vector and the transformation matrix correspondingly increase. cubic. or other terms are included. The transformation matrix A is obtained by minimizing the weighted square error given by E¼ N X i¼1 (6:25) Wi kyi À Aei k2 x (6:26) The above expression for E can be expanded as E¼ N X i¼1 Wi ðyi À Aei Þðyi À Aei ÞT ¼ x x N X i¼1 Wi yi yT À 2A i N X i¼1 Wiei yT þ A x i N X i¼1 WieieT AT x xi (6:27) Differentiating the above equation with respect to A and setting it equal to zero yields N N X X qE ¼ À2 Wi yieT þ 2A WieieT ¼ 0 xi x xi qA i¼1 i¼1 (6:28) . It is based on weighted least-square regression.2 MOVING-MATRIX INTERPOLATION The moving-matrix approach is another nonlinear technique for interpolation of irregularly spaced or scattered multidimensional data [5].Interpolation of Multidimensional Functions Therefore. The interpolated value ^ at point x is given as y ^ ¼ Ae y x where e ¼ [x 1]T is the augmented vector x x A is the transformation matrix of size M Â (M þ 1) If quadratic.3. [C M Y] ¼ [98:27 161:53 176:64] 267 PN P27 P27 Àm À2 CMY k[50 80 À 14] À L*a*bi *kÀ2 i¼1 y di i¼1 y d ^ ¼ PN i Àm ¼ P27 i i ¼ i¼1 27 i (b) y P À2 À2 i¼1 di i¼1 k[50 80 À 14] À L*a*bi *k i¼1 di ¼ [27:35 Therefore.

SOLUTION Wi ¼ 1 1 1 ¼ ¼ dim þ e jx À xi j2 þ 0:0001 j0:5 À xi j2 þ 0:0001 i ¼ 1.34 À0.0266 0. and compute the corresponding value of P(x) at x ¼ 0.7 Consider the 1-D function y ¼ P(x) given by the LUT shown in Table 6.279 0.55 0. These two parameters affect the locality of the regression. We now consider a sample example.07 0. .27 0.66 0.9 One-Dimensional Function LUT x y 0 0.545 0.80 0.461 0.702 1 1 . 10 TABLE 6.0334 0. Use m ¼ 2 and e ¼ 10À4.523 0.5. Example 6. .0079 0.41 0. which means that only points closest to the desired point will be considered as signiﬁcant points. which gives more weight to the points closer to the desired point and less weight to points that are farther away from the desired point.28 is used to solve for A. It is inversely proportional to the distance. . The variable parameters of this algorithm are m and e. interpolate.89 0. Large values of m and small values of e give Wi more local behavior.9. The weight Wi is given by Wi ¼ 1 dim þ e (6:30) where di ¼ kx À xik is the Euclidean norm distance.419 0.268 Control of Color Imaging Systems: Analysis and Design Equation 6. 2. Use the moving-matrix approach. The result is A ¼ SPÀ1 where S¼ N X i¼1 (6:29) Wi yieT xi WieieT x xi P¼ N X i¼1 The weight Wi is a function of the distance from the input point x to all the points xi in the LUT. .

where a2 ) 0 and I is the 3 Â 3 identity matrix. The transformation matrix A is given by A ¼ SP À1 ¼ N X i¼1 !" Wi yieT xi N X i¼1 #À1 WieieT x xi (6:31) As can be seen.1.97 2 29. Let B(0) ¼ a2I and k ¼ 0.10 Weights as a Function of i i Wi 1 15. Choose e to be a small positive number.02 4 1323.25 can be implemented using recursive least square (RLS).4.3. it is possible to compute the inverse of P without using any matrix inversion.2). By using the RLS algorithm (Section 7.Interpolation of Multidimensional Functions 269 TABLE 6.3 RECURSIVE LEAST-SQUARE IMPLEMENTATION OF MOVING-MATRIX ALGORITHM The moving-matrix interpolation algorithm given by Equation 6. Step 3: Compute w(k) ¼ ke À ek k þ e x x (6:32) .7 7 1323.87 5 6038. we would like to compute A without i¼1 using matrix inversion. Since the P x xi matrix P ¼ N WieieT may be ill conditioned. This can also be achieved by recursive computation of A using the following algorithm: Step 1: Initialization.28 6 9411.10. matrix inversion is required to compute the matrix A.97 The weights for different values of i are shown in Table 6.16 3 345.95 9 43.87 8 121.04 10 15. S¼ N X i¼1 Wi yi xT ¼ i 10 X i¼1 Wi yi ½xi 1 ¼ ½83:8135 144:9269 3 2 10 10 P P Wi x2 Wi xi 7 ! i N 10 6 X X 176:0015 324:6467 i¼1 7 6 i¼1 P¼ WieieT ¼ Wi ½xi 1T ½xi 1 ¼ 6 x xi 7¼ 10 10 5 4P P 324:6467 634:3268 i¼1 i¼1 Wi xi Wi i¼1 i¼1 A ¼ SPÀ1 ¼ [0:9789 À 0:2725] ! 0:5 ^ y ¼ Ae ¼ [0:9789 À 0:2725] x ¼ 0:2169 1 6. Step 2: Change k ! k þ 1.

14. deﬁned mathematically as L*a*b* ! CMY. theoretically speaking.4. The reversal can be easily obtained by swapping the data from the forward LUT. structured input grid.. For the purpose of this discussion. The inverse ^ printer map PÀ1 is a three-to-three map. we need to build the inverse map of the forward LUT [6. conjugate gradient (CG).e.7]. and iteratively clustered interpolation (ICI) algorithms. It may give multivalued outputs. However. then.4 LOOKUP TABLE INVERSE 6.5.1). . While generating the printer forward table. The CMY values are converted to CMYK using gray-component replacement or under-color removal (GCR=UCR) algorithms before printing on a CMYK to L*a*b* printer (Section 7. 6. Also. otherwise. for colors at the gamut boundary. an inverse LUT is just the reversal of the forward LUT.2 INVERSE PRINTER MAP Let us describe the inversion process using the CMY to L*a*b*.1 INTRODUCTION A color printer can be seen as a device that is mapping a requested color in the image into a device independent color (or printed color). An approximation to this color mapping can be obtained by constructing a forward map of the printer using experimental data in the form of a LUT. The resulting inverse LUT will not conform to the structured speciﬁcations required for the input nodes. such as tetrahedral. as shown in Figure 6. P. if we select node colors appropriately in the forward LUT. stop and compute matrix A using the following equation: A ¼ D(N)B(N) (6:34) (6:33) 6. To reproduce colors accurately. the GCR=UCR function is embedded inside the printer map. Multidimensional interpolation approaches. Let PÀ1 denote an estimate of the printer inverse map. and only in-gamut colors are considered. it is important to note that the various other approaches described in Chapter 7 do not always use these methods. a three-to-three printer ^ forward map.270 Control of Color Imaging Systems: Analysis and Design and C(k) ¼ B(k)ek x b(k) ¼ w(k) þ eT C(k) xk B(k þ 1) ¼ B(k) À C(k)C T (k)=b(k) D(k þ 1) ¼ D(k) À ek yT =w(k) x k Step 4: Is k ¼ N À 1? If the answer is no. are often used to restructure the inverse LUT when such LUTs are constructed with experimental data such that the inverse LUT ﬁnally ends up with a uniformly sampled. P. go to Step 2. this type of inverse LUT may not be well deﬁned. Each entry in the inverse LUT is called a node. One form of this LUT associates the input colors in L*a*b* space to printer speciﬁc CMYK space. It is desirable to have an inverse LUT with input nodes regularly spaced on a sequential plane (i. More description about the inverse LUTs can be found in Chapter 7. a structured input).4.

such that the ^ composition PÀ1[P(Á)] is as close as possible to the identity matrix in the minimum mean-square error (MMSE) sense. and À127 b 128. is required so that the structured inverse conforms to the performance index deﬁned by Equation 6. x] ¼ 1kP(y) À xk2 ¼ 1[P(y) À x]T [P(y) À x] 2 2 (6:35) Notice that we do not have an exact model for P. 2. When this happens. (3) Obtain the corresponding value of yi ¼ [Ci Mi Yi] by minimizing the cost function given by Equation 6. as shown schematically in Figure 6.14. Another interpretation for the inverse is that. A simple cost function to be minimized can be * formulated using the DEab color difference formula between the requested input * color x ¼ L*a*b* and the printed output color z ¼ P(y) ¼ L*a*bout : in E(y) ¼ DE[P(y). The interpolation methodology used to solve the threeto-three inverse problem is summarized as follows: (1) Obtain the forward LUT P for the given color printer. That is. we grid the CMY color space at the input of the GCR algorithm.14 Forward and inverse printer maps. The ^ problem of interest is to compute a structured inverse LUT. a CMY to L*a*b* printer whose structured inverse. The printed color zj for each grid node yj is obtained from experiments on the actual printer. . simply. L*a*b* to CMY. in ^ we seek an inverse printer map PÀ1 such that the printed color z and the target x are as close as possible (or same).35. i ¼ 1. we can also restrict the printer forward map without the GCR algorithms. . where input L*a*b* points are on a 3-D grid of size h Â h Â h having a dynamic range 0 L 100. we can uniformly sample the CIELab color space. only an approximation given by the forward LUT is available. (2) Select the collection of target colors xi. . which is created to compensate for the nonlinearity in P. We assume that every grid point in the sampled CIELab color space is either inside the printer gamut or it has been mapped to a point inside the gamut by an appropriate gamut mapping algorithm (Section 7.35. For this. For this discussion. where * * * xi ¼ ½ Lix aix bix T . To achieve this. PÀ1.6). This inverse printer map PÀ1 is a key component of many color control algorithms used in digital printers. we say there is a colorimetric ^ match between the target color and the printed color. .Interpolation of Multidimensional Functions 271 x L*a*b* in Printer inverse map y CMY Forward printer map z L*a*b* out FIGURE 6. Process the CMY grids through the GCR algorithm to create print ready CMYK values at each CMY grid node. h. An approach to solve the three-to-three optimization problem is described next. . for a given target color x ¼ L*a*b*. À127 a 128.

otherwise go to Step 2.4. * The parameter e is the minimum required DEab and can be selected to be any arbitrary small number. Step 3: Let k ¼ k þ 1. .272 Control of Color Imaging Systems: Analysis and Design 6. then stop. and kmax (their meanings and selections will be discussed shortly).36 results in T y(k þ 1) ¼ y(k) À mJk {P[y(k)] À x} qE qy(k) (6:36) (6:37) (6:38) The term Jk is the 3 Â 3 Jacobian matrix computed at the kth iteration. If E[y(k)] < e or k > kmax. The index kmax is the maximum number of iterations.3 ITERATIVELY CLUSTERED INTERPOLATION The ICI algorithm is a gradient-based optimization method. with the initial point for the optimization generated through an iterative technique [8]. e. The step size m controls the rate of convergence and should be selected to achieve fast algorithm convergence and meet accuracy requirements.37 into Equation 6. Step 2: Update y(k) using the recursion formula y(k) ¼ y(k þ 1) À m where qE T ¼ 2Jk {P[y(k)] À x} qy(k) Inserting Equation 6. The upper bound for parameter m is derived in the next section. It is given by qP½y1 (k) 6 qC(k) 6 6 6 qP½y2 (k) Jk ¼ 6 6 qC(k) 6 6 4 qP½y3 (k) qC(k) 2 3 qP½y1 (k) qP½y1 (k) qM(k) qY(k) 7 7 7 qP½y2 (k) qP½y2 (k) 7 7 qM(k) qY(k) 7 7 7 qP½y3 (k) qP½y3 (k) 5 qM(k) qY(k) (6:39) The Jacobian matrix Jk is computed numerically at each iteration using the forward printer map. Let k ¼ 0 and assume the initial condition Ä Å y(0) ¼ Cy (0) My (0) Yy (0) . The ICI algorithm is implemented in the following steps: Step 1: Select algorithm parameters m. The * algorithm will stop either when DEab is less than m or when the number of iterations reaches kmax.

we have T e(k þ 1) À e(k) % ÀmJk Jk e(k) (6:41) (6:42) (6:43) or À Á T e(k þ 1) ¼ I À mJk Jk e(k) (6:44) The Jacobian matrix Jk is not changing rapidly from one iteration to another.Interpolation of Multidimensional Functions 273 6. 2. 3 (6:46) Therefore.4.38.1 Selection of Step Size Parameter m The updating equation for y(k) is given by T y(k þ 1) ¼ y(k) À m Jk {P[y(k)] À x} (6:40) Deﬁne the error e(k) as e(k) ¼ P[y(k)] À x Use the ﬁrst-order linear approximation e(k þ 1) À e(k) ¼ P[y(k þ 1)] À P[y(k)] % Jk [y(k þ 1) À y(k)] Using the updating law given by Equation 6. À Á T e(k þ 1) ¼ I À mJ0 J0 e(k) (6:45) Therefore. a more conservative. À Á li I À mJ0 J T < 1 0 i ¼ 1. method Since lmax J0 J0 for obtaining an upper bound for parameter m is 0<m< 2 À TÁ Tr J0 J0 (6:48) . for the iterations to converge. T This requires that all eigenvalues of matrix I À mJ0 J0 must lie inside the unit circle in the complex plane. that is. but easier to compute.3. the parameter m should be selected to satisfy the following condition: 0<m< 2 À TÁ lmax J0 J0 (6:47) À TÁ À TÁ Tr J0 J0 . With this assumption. therefore. we assume that Jk ¼ J0. the error e(k) must approach zero as k ! 1.

To ﬁnd the inverse of z ¼ L*a*b* using the tetrahedral interpolation method. Choose the initial condition to be y0.4. yaux2. .2 Algorithm Initialization To ﬁnd an initial estimate y(0) ¼ [C(0) M(0) Y(0)]T for the algorithm. . . and Y axes around yaux. . such that P(yaux) ¼ zaux. Assume a set of CMY nodes (yi) and their corresponding L*a*b* nodes (zi) are given in the form of a LUT. N. For this reason. Use the LUT P and trilinear interpolation to map these N points to obtain the corresponding values in L*a*b* color space. . .4 TETRAHEDRAL TECHNIQUE The tetrahedral interpolation technique covered in Section 6. where z0 ¼ P(y0). M. since zaux is a grid point in the forward LUT. we follow the following steps: * * * T (1) Find the auxiliary point zaux ¼ ½Laux aaux baux in the forward printer LUT that is closest to the target color x in an L2 norm (Euclidean distance) sense. . .2. . (4) Denote by z0 the closest point to the set zauxi i ¼ 1. Step 2: Given a target x ¼ L*a*b*. yauxN by moving along the C.274 Control of Color Imaging Systems: Analysis and Design 6.3. ﬁnd the tetrahedral that contains the z vector. Call them zauxi ¼ P(yauxi) i ¼ 1. Let {z1 z2 z3 z4} denote the vertices of the tetrahedron that z belongs to in L*a*b* and let {y1 y2 y3 y4} denote their corresponding vertices in CMY color space.4. . . 2. that is. 2. . (3) Select N grid points in a neighborhood of yaux. Step 3: Compute the corresponding CMY value using the following equation: CMY ¼ y ¼ ACMY AÀ1 ðz À z1 Þ þ y1 Lab where ACMY ¼ ½y2 À y1 and ALab ¼ ½z2 À z1 z3 À z1 z4 À z1 (6:51) y3 À y1 y4 À y1 (6:50) (6:49) . Its low computational complexity makes it an affordable approach for constructing an inverse LUT [2]. (2) Find the point yaux. 6.3 can be used to ﬁnd the inverse printer map. generate a cluster of N points yaux1. we perform the following steps: Step 1: Partition the CMY color space into tetrahedral segments. it is a widely used methodology for inverse computation. N. Note that yaux can be found easily.

Assume a forward printer map z ¼ P(y). respectively. The goal * is to ﬁnd y for a given z ¼ x by minimizing the DEab error between x and P(y). where y 2 R3 is a CMY color patch and z 2 R3 is its corresponding L*a*b* value.4. min E(y) ¼ min DE[P(y). The accuracy of the CG is similar to the ICI algorithm. x] ¼ min 1kP(y) À xk2 2 y y y ¼ min 1[P(y) y 2 À x] [P(y) À x] T (6:52) The iteration to obtain the solution using CG method is given by y(k þ 1) ¼ y(k) þ ak dk (6:53) where dk and ak are the search direction vectors and the step size at the kth iteration. . DEab . The iteration stops when the MSE. The initialization algorithm to obtain an initial estimate y(0) is similar to the technique used for ICI. is less than e or the number of iterations exceeds Kmax. that is. where matrix A is symmetric and positive deﬁnite [9].5 CONJUGATE GRADIENT APPROACH The CG technique is an iterative technique used in solving constrained optimization problems and systems of linear equations of the form Ax ¼ b. the computational complexity of the CG method far exceeds the ICI approach.39 The updating equations for the search direction and the step size are given as ak ¼ À T dk Jk eT (k) T T dk J k J k dk (6:54) (6:55) (6:56) (6:57) (6:58) gkþ1 ¼ e(k þ 1)Jkþ1 bk ¼ gT ðgkþ1 À gk Þ kþ1 T dk ðgkþ1 À gk Þ T dkþ1 ¼ ÀgT þ bk dk kþ1 The criteria for stopping the iterations are given by the threshold parameter e and the * maximum number of iterations Kmax. however. The CG algorithm can also be used to solve unconstrained optimization problems such as ﬁnding the inverse printer model described in the previous section. The initial search direction is given by dk ¼ Àgk ¼ Àe(k)Jk where e(k) ¼ P(y(k)) À x(k) Jk is the Jacobian matrix deﬁned by Equation 6.Interpolation of Multidimensional Functions 275 6.

807 5.4. The LUT is a mapping from a device-dependent CMY to a device-independent L*a*b* color space.1182 * Mean DEab þ 2s 2.15).923 5. and (d) CG.002 * Maximum DEab 3. The 240 in-gamut colors are provided as input to the inverse LUT and the output of the inverse LUT is interpolated through the forward LUT. The interpolation in the forward LUT uses a trilinear algorithm. (b) moving matrix. The error results are shown in Table 6. a 133 printer forward LUT size was chosen. (c) ICI.044 Standard Deviation 0. which is common to all methods. We modeled a legacy Xerox color digital printer by creating a 133 LUT.) comparing inversion algorithms.151 0. FIGURE 6.11. the DEab statistics are simulated. * The quantitative metric used to evaluate the inversion accuracy is the DEab accuracy for interpolated colors between the mapped input L*a*b* and the resulting * output L*a*b*.276 Control of Color Imaging Systems: Analysis and Design OF 6.809 0.149 0. The output of the interpolation is compared through the input to get the error performance.681 . View of 240 test colors used for TABLE 6. For 240 in-gamut test colors (Figure 6. The inverse map is generated by uniformly sampling the L*a*b* space and ﬁnding the inverse of each L*a*b* node on the grid using one of the following algorithms: (a) Shepard.5448 0.15 (See color insert following page 428. The grid points in the CMY color space are uniformly sampled.023 0. Out-of-gamut colors were ﬁrst mapped to the nearest point in the gamut by a technique that maintains the hue angle.5243 0. For this simulation.764 1.2812 * Minimum DEab 0.11 Comparison of Four Different Printer Inverse Algorithms Algorithm Shepard Moving matrix CG ICI * Mean DEab 1.236 0.7346 2.187 0.001 0.6 COMPARISON DIFFERENT INVERSION ALGORITHMS We now brieﬂy compare four different mapping algorithms when used for building the inverse map.644 1.113 1.

which are now in the form of a CMY N×N×N P Lab(out1) ˆ P Upsample to N×N×N LUT Downsample N1 × N1 × N1 LUT + _ Compute * ΔEa b * ΔEa b Lab(out2) FIGURE 6.16 Optimum color selection method illustrated for obtaining minimal colors for measurement of a three-to-three printer forward map. An important performance measure of the inverse LUT is the accuracy of inversion. as indicated above. To compensate for the printer drift. Hence. . the inverse map is a LUT that associates points in the printer output color space to points in the printer input color space (L*a*b* ! CMYK). * The average DEab between the original L*a*b* and L*a*b1 * (the upsampled value from the downsampled LUT corresponding to the patch represented by L*a*b*) is deﬁned as * DEab ¼ K 1 X kL*a*b* À L*a*b1 *(i)k K i¼1 (6:59) * These critical colors are selected by minimizing DEab. when we * upsample the downsampled LUT to the original size.5.g. often the manufacturers increase the LUT size greatly (e.5 COMPRESSION OF LOOKUP TABLES 6. It is obtained simply by ﬁrst printing and measuring these selected critical colors. the updating of the three-to-three inverse printer map is straightforward. it is highly desirable to update the inverse printer LUT on a regular basis. then this requires printing and measuring a large number of color patches. Once these critical colors are identiﬁed. The process is shown schematically in a block diagram in Figure 6. N may be chosen equal to 33 when compared to 17). To achieve an improved accuracy for colors inside the printer gamut.. and then upsampling the measured L*a*b* values.Interpolation of Multidimensional Functions 277 6. If we have to update each node in the inverse LUT through measurements. the average DEab between the original L*a*b* and the upsampled L*a*b* is minimized.16. for an N Â N Â N LUT. a better way to do this is by downsampling the input color space from N Â N Â N to M Â M Â M (M < N) by selecting the critical colors for measurement such that.1 INTRODUCTION In general.

to the full size and then applying an inverse algorithm on the upsampled LUT to obtain the full-size inverse map. There are different techniques available in the literature for selecting these critical colors [3.10]. In the following sections. . The MSE is given by f x E¼ M 1 X À Á f x j À Lk ð x i Þ 2 M i¼1 (6:63) . The DO algorithm is based on dynamic programming. Interested readers can refer to Refs. 6.5. These algorithms are very complex and their derivations are not given in this book. We would like to choose N < M points ^ ¼ [^1 ^2 Á Á Á ^N] such that ^1 ¼ x1. The 3-D approach will be used to ﬁnd the critical colors for measurement of a three-to-three printer forward map.2 DOWNSAMPLING USING SEQUENTIAL LINEAR INTERPOLATION SLI is a numerical technique to optimally downsample a uniformly spaced LUT to any desired size.3 DYNAMIC OPTIMIZATION ALGORITHM The DO algorithm selects a ﬁnite number of colors (or points in L*a*b color space) * by minimizing the MSE (DEab) between the actual printer output and the upsampled printer output constructed using a ﬁnite number of points as deﬁned by * DEab ¼ kL*a*b*(out1) À L*a*b*(out2)k Lab(out1) ¼ P(CMY) ^ Lab(out2) ¼ P(CMY) (6:60) (6:61) (6:62) ^ The LUT (P) is obtained by upsampling the smaller LUT containing the ﬁnite number of critical colors. we ﬁrst outline the 1-D case and then extend the approach to two and three dimensions. This is shown in Figure 6.5. 6. Piecewise linear homeomorphism is another technique illustrated in Ref. x2. xM. ^N ¼ xM.16. where x takes M discrete values x1. its inverse LUT (inverse of P) can be computed using ICI or the other algorithms mentioned above. Once the upsampled forward ^ ^ LUT (P) is constructed for a three-to-three map. To illustrate the algorithm in detail.278 Control of Color Imaging Systems: Analysis and Design reduced-sized forward LUT. x x x x x x and ^ & x while minimizing the MSE resulting from the piecewise linear approxix mation deﬁned by ^(^) and f(x) over x. Numerical examples are provided for comparison of these two approaches. 6. The two most promising techniques for downsampling of multidimensional LUTs are SLI and dynamic optimization (DO) algorithms. . [10].3. which uses a multistage decision process and the * performance criteria such as minimization of DEab error criteria. . [3.10.1 One-Dimensional DO Algorithm Consider the 1-D discrete function f(x). we consider both SLI and DO algorithms. .11].5.

x x the number of grid points has to be greater than or equal to three. and error. In this array. N ! 3. These indices and their corresponding MSEs are required for the two-stage optimal search. In the two-stage search. We call this the 1-D single-stage grid allocation algorithm. Let the solution be xj. We repeat this process for x2 through xN À 2 and form the array of numbers as shown in column 1 of Table 6. and assign these two numbers to x1 and write it as {j1. we need to minimize . E1}. The MMSE is given by j M 1 X 1 X j f ðxi Þ À L1 ðxi Þj2 þ j f ðxi Þ À L2 ðxi Þj2 M i¼1 M i¼jþ1 E* ¼ (6:65) where L1(x) and L2(x) are À Á f xj À f ð x1 Þ L1 (x) ¼ ð x À x1 Þ þ f ð x1 Þ xj À x1 À Á Á À Á f ð xM Þ À f xj À L2 (x) ¼ x À x j þ f xj xM À x j E* and the index j are found by the following optimization problem: j ¼ arg mink [E(k)] ¼ arg mink k M 1 X 1 X j f ðxi Þ À L1 ðxi Þj2 þ j f ðxi Þ À L2 ðxi Þj2 M i¼1 M i¼kþ1 (6:66) (6:67) (6:68) (6:69) E* ¼ E(j) This means that if we start from x1 and wish to locate one grid point between x1 and xM to approximate f(x). Since we always need the ﬁrst and the last points for interpolation. that point will be xj and the corresponding MMSE will be E*. Using dynamic programming.12. which means that we need to ﬁnd one grid point ^2 between x1 and xM x such that the total MSE given by Equation 6. where 1 < j < N. we try to locate two optimal grid points between xi and xN such that the total MSE is minimized.Interpolation of Multidimensional Functions 279 where Lk(x) is the straight-line segment joining f(^k) to f(^kþ1) and is given by x x Lk (x) ¼ f ð^kþ1 Þ À f ð^k Þ x x ð x À ^k Þ þ f ð^k Þ x x ^kþ1 À ^k x x (6:64) for ^k x ^kþ1. Now assume that N ¼ 3. E1 ¼ E*. j1 ¼ j. We deﬁne a new index. Ei is the MSE between the original function and the linearly interpolated function using grid points bxi xj xNc over the closed interval of [xi xN].63 is minimized.

6. 2. . . . E1] [j2. C. E2] [j3. degrees of freedom are less than N Â N. We would like to choose N Â N grid points out of M Â M available data points such that the mean-square interpolation error between the original function and the approximation obtained by upsampling the N Â N LUT is minimized. The solution is obtained by direct optimal search and is denoted by the two indices ki and lj with its associated four boundary points and MMSE J1(i.12 One-Dimensional DO Algorithm Point x1 x2 x3 . Since the boundary points should be included for interpolation. xm] Then the optimum two-point grid including boundary points will be [x1 xk xl xN]. assume we start with point [xi. we can now build an n-stage optimization process for the 1-D case with N optimal grid points [x1 xm1 xn1 Á Á Á xN1]. EE2] [k3. as shown in column 2 of Table 6. j. We call this the 1-D two-stage grid allocation algorithm. Using the same approach. 0] — — J¼2 [k1. where X ¼ [x. j) ¼ Ei. . if A is an internal optimum point. j ¼ 1.5. . li ¼ jki (6:70) where Em is the MMSE corresponding to j ¼ 1 Eim is the MSE between the original function and the linearly interpolated function using grid points [xi.17) should also be included in this set. xNÀ4 xNÀ3 xNÀ2 xNÀ1 xN J¼1 [j1.280 Control of Color Imaging Systems: Analysis and Design TABLE 6. ÁÁÁ — — — — J¼N À2 [m1. . . n1. EENÀ4] [NÀ2. . M (6:71) . E3] . ENÀ3] [NÀ1.2 Two-Dimensional DO Algorithm Consider the 2-D discrete function f(X). l1. D. . .12. EE3] . ÁÁÁ — — — — k ¼ arg minm ½Eim þ Em . xm] over the interval of [xi. . l2. . . 0] — — — J¼ÁÁÁ ÁÁÁ ÁÁÁ ÁÁÁ . lNÀ4. m ! i þ 1 EEi ¼ Eiki þ Eki . . . In the single-stage 2-D algorithm. as shown in Figure 6. . [kNÀ4. . For example. l1. then the four boundary points B. EE1] [k2. With three indices given by ½ ki lj Ei. ENÀ4] [jNÀ3. . NÀ1. yj] and try to optimally select one internal grid point in the set extending from this point to the boundaries of the function f(X). . This process is repeated for each point in the set associated with the domain of the underlying function. The 2-D DO algorithm is very similar to the 1-D case. [jNÀ4. . y] takes M Â M grid points uniformly spaced in the x–y plane.j i. .3. .17. ETotal] — — . and E (as shown in Figure 6.

( j. then the total MSE is E ¼ Eði. Let the solution be i ¼ ni.mÞ þ E1 (n. j ¼ mj.3. assume that we start from point {i j k} and form a cube extending this point to the boundaries of the underlying function. we will try to select two internal points in the same set.m) is the MSE resulting from interpolating all the grid points between indices i to n and j to m using all necessary boundary points. Then the indices ni. which is an extension of the 2-D algorithm.mÞ þ En. m correspond to the two optimum internal points (total of 14 points including boundary points) with the MSE given by E2 (i. it is a multistage decision process described by the following steps. j). ki . lj A E D xj xki xM B x FIGURE 6. In the single-stage decision process. The 3-D DO algorithm.m (6:72) where E(i. The goal is to ﬁnd one point inside the cube with its associated boundary points such that the MSE between the . where X is a 3-D vector. We now minimize E given by Equation 6.ð j.17 Grid points in x–y plane.Interpolation of Multidimensional Functions y C yM yl j 281 yj Ei.jÞ. In the second stage.5. 6. Similar to 2-D cases. If the indices corresponding to the ﬁrst optimal point are n and m. can be used for optimally selecting these grid points.3 Three-Dimensional DO Algorithm (6:73) Consider a 3-D discrete function f(X) represented by a set of pairs (X.ðmj . These points are uniformly spaced in the support space of f(X) forming an M Â M Â M LUT. f(X)). mj. j . n.mÞ þ En.ðmj . m) This process is continued until we select N Â N grid points.m ¼ Eðni nÞ. The goal is to downsample this 3-D LUT to a smaller LUT of size N Â N Â N while minimizing the MSE between the original and the upsampled LUT.72 with respect to i and j. j) ¼ Eðni nÞ.

mj. Let the optimal indices be li. and DO algorithms to select 5.4 0.18 algorithm. m1j . we minimize E2 with respect to the indices l1i. m1j. k) is the interpolation error between the original and the interpolated functions with the cube extending from grid point {i j k} to grid points l1i. and n1k. the DO algorithm outperforms the other two algorithms. and n1k. n1k ¼ E(i. . m1j.8 Consider the 1-D nonlinear function shown in Figure 6. and nk with an MSE of Eijk. similar to 1-D and 2-D algorithms.7 0. As can be seen. . m1j. We then use linear interpolation and interpolate to the original size of the function and compute the MSE between the original function and the upsampled version of the downsampled function.6 0. 6. Nonlinear 1-D function with eight grid points optimally selected by the DO . Nonlinear function 1 0. n1k.5 x 0. Now.9 1 FIGURE 6.19 shows the MSE as a function of grid size for three different algorithms (uniformly sampling with reduced samples and SLI). We continue this process until we ﬁnd the solution to the N-stage problem. we obtain the solution to all stages up to and including N.2 0 0 0.1 0.18.8 0. In the second stage. .4 0. If the solution for the ﬁrst point has indices l1i. Figure 6.2 0. Example 6.282 Control of Color Imaging Systems: Analysis and Design interpolated and original grid point values is minimized. then the total MSE is given by À Á E2 l1i .3 0. Note that while ﬁnding the solution to the N-stage grid allocation problem. Note that the solution to the ﬁrst stage is used for ﬁnding the optimal solution in the second stage. j. k) þ El1i m1j n1k (6:74) where E(i. j. SLI.8 f (x) 0. We apply uniform sampling. 20 grid points. .6 0. we need to ﬁnd two grid points within the same cube in order to minimize the MSE as in stage one.

902 6.946 3.8 0. y) x 0 0 0 0 0.47 2.2 ΔE 1 0.20.8 1.4 0.13.4 0.2 0 0.4 0.8 1.8 0.2 0 0.0869.92 5.32 1.2 1.9 Consider a discrete 2-D function deﬁned by the 2-D LUT of size 4 Â 4 as given in Table 6.4 1.8 1.2 1.4 0.4 0.2 f 1 1.4 0.4 0.6 1.48 2. The resulting MMSE is 1.206 3.134 1.646 1.678 3.Interpolation of Multidimensional Functions 2 1.8 1.2 y 0 0.13 Two-Dimensional Function f(x.182 5.018 7.2 0 5 10 Grid size N 15 20 Uniform SLI DO 283 FIGURE 6.994 . Example 6. The optimal grid points (9 points) are shown in Figure 6.002 5.4 0.6 0. We downsample the LUT to 3 Â 3 using the 2-D DO algorithm.558 1.8 1.2 1.8 0.19 MSE as a function of grid size.4 0.8 1.8 0. TABLE 6.2 0 0.

. This LUT is then downsampled to 3.21 The system performance is measured in terms of the error between the mapped input and output L*a*b* values: * E ¼ L*a*bout À Labin (mapped) (6:75) The forward printer model P is constructed by printing 133 uniformly sampled. . generating a LUT of size 5 Â 5 Â 5 ¼ 125 samples.15. Example 6. . 4. y 1 The function f(x. y. Example 6. z) ¼ x þ x2 þ xy þ 1:5y1:6 þ xy4 þ 0:1 0 x. 13 cubes and then upsampled to the original size (13 cubes).8 1.1107 and the optimal grid points are shown in Table 6. z) is uniformly sampled with Dx ¼ Dy ¼ Dz ¼ 0.2 0.4 0.10 Consider the following 3-D nonlinear function: f (x.76 is achievable with the reduced measurements of size 125 optimal colors. The simulation results for a commercial Xerox printer are shown in Table 6. a mean DE2000 ¼ 0. As can be seen.2 FIGURE 6. The 3-D DO algorithm is used to downsample the LUT to a smaller LUT of size 3 Â 3 Â 3 ¼ 27 samples.20 Grid locations for 2-D example. . . The 240 test patches that are used to evaluate the overall system performance consist of 216 patches inside the printer gamut and 24 patches inside and on the neutral axis (a* ¼ b* ¼ 0) of the printer gamut.25 over a cube of size [0 1]Â[0 1]Â[0 1].284 Control of Color Imaging Systems: Analysis and Design 1.2 0 0 0. y. Applying ICI to the newly formed upsampled LUT forms the three-to-three inverse LUT.2 0.4 0. The resulting MMSE is 1.8 0.11 The overall system for constructing a three-to-three inverse LUT with reduced patch measurements is shown in Figure 6.14. CMY color patches and measuring their L*a*b* values using a spectrophotometer.

008 1.094 4.6 Example 6.5 1 1 1 1 1 1 1 1 1 y 0 0 0 0.21 depicts a chain of events typical for a printing system image path. Transformation T2 moves the image from ROMMRGB to the device-independent color space L*a*b*.5 0.753 2. Note that the transformations from sRGB to ROMMRGB and back to sRGB are nonlinear.75 1 0 0.008 3. The results of the simulation for a 512 Â 512 Lena image using a Xerox Phasor 770 digital printer are shown in Table 6.1 0.5 0.75 1 0 0.094 2.5 0.75 1 f 0.594 0.5 0.35 2. Color space ROMMRGB has been developed by Eastman Kodak Company as a standard color space for the storage and manipulation of digital images. The input image is in a device-dependent color space such as sRGB.5 1 1 1 0 0 0 0.75 1 0 0.85 3.14 Result of Downsampling x 0 0 0 0 0 0 0 0 0 0.416 3.594 1.5 1 1 1 z 0 0.5 0. .594 0.6 0.094 3.75 1 0 0.5 0.6 1.5 0.411 4.75 1 0 0.35 1.916 5.75 1 0 0.1 0.5 0.Interpolation of Multidimensional Functions 285 TABLE 6.75 1 0 0.5 0.5 0.16.5 0.12 Figure 6.85 1.60 4.594 1. we will assume it is a standard 24-bit color image that is transformed (T1) [12] to reference output medium metric ROMMRGB.6 1.1 0.75 1 0 0.5 0.5 0. For the present purpose.1 2.5 1 1 1 0 0 0 0.5 0.1 3.

6.6. The second consideration is the smoothness of the ﬁt. then its second derivative must . When ﬁtting a function to data.21 Printing system image path.286 Control of Color Imaging Systems: Analysis and Design sRGBin T1 ROMMRGBin T2 Labin Gamut mapping Mapped Lab CMYin ˆ P –1 CMYK printer P ˆ Compute P –1 using ICI UCR/GCR P1 Labout Downsample P using dynamic optimization (DO) ~ P ~ Upsample P using trilinear interpolation For visualization on a typical display screen sRGBout T1 –1 ROMMRGBout T2 –1 FIGURE 6. Ideally. The L2 norm (MSE) is a measure normally used for this purpose. The smoothness of the ﬁt is measured in terms of the ﬁrst or second derivative of the underlying function. there are two important considerations. The ﬁrst is the goodness of the ﬁt.10. the ﬁt and the data would have the same value at every point.1). Other norms such as L1 and L1 can also be used (Section 3. The goodness of the ﬁt is typically measured by an error metric deﬁned in terms of the difference between the available data and the ﬁt.1 INTRODUCTION Data ﬁtting is an important task in printer calibration and characterization.6 SMOOTHING ALGORITHM FOR MULTIDIMENSIONAL FUNCTIONS 6. If the function is smooth.

29 3.65 0.38 7.11 7. As an application. To ﬁnd the best ﬁt.6.70 1.92 4.69 3.56 6.15].41 4.56 3. Finally.41 4.61 3. We use the second derivative as a measure of the smoothness of ﬁt. On the other hand.18 Mean DE 3.57 95th Percentile DE 4.94 8.6.96 1.90 4.12 1. it can be applied to the 2-D printer calibration [16].24 0. In the 1-D case.0 3.94 TABLE 6.42 7. we use the 3-D or 4-D proﬁling systems. This will create problems if the data is noisy.46 1. and for the 2-D case. then it might not approximate the data very well.08 7.2 presents the 1-D through 4-D smoothing algorithm.09 1.20 0.3 describes applications to printer calibrations with simulation results.19 2.74 1.48 3. Section 6. it can be applied to the gray balance calibration [14.30 4. If no restriction is made on how smooth the ﬁt must be. the proposed algorithm can be applied to digital printer calibrations.18 1.80 * 95th Percentile DEab 4.53 4. both of these factors must be considered.14 7.1 0.12 be small. if the ﬁt is too smooth.93 3.27 1.05 3.76 1.16 DE Statistics for Different LUT Sizes Downsample LUT Size 27 64 125 216 2197 * Mean DEab 3. we consider the problem of smooth curve ﬁtting in 1-D [13] through 4-D cases.00 1. as a 3-D example.07 3.Interpolation of Multidimensional Functions 287 TABLE 6. In this section. since we must not ﬁt to the noise in the data.37 0.84 8.99 0.77 0.78 7.51 3.11 2.45 3. Section 6.0 3.08 3.05 8.03 * Maximum DEab 8.37 4. .15 DE Statistics for Downsampled LUTs of Different Sizes Reduced Measurement Colors 27 48 125 216 343 512 4913 Error Metric * DEab DE2000 * DEab DE2000 * DEab DE2000 * DEab DE2000 * DEab DE2000 * DEab DE2000 * DEab DE2000 Maximum DE 8. then it ends up matching the data exactly.13 7.17 2.

.6. . 7 Cn ¼ 6 . . . is used to measure the smoothness of the ﬁt f.7 6 0 1 À2 1 0 ÁÁÁ . 7 6 6 . 1 À2 1 0 7 7 6. J1 ¼ nÀ1 XÀ i¼0 Á2 fi À ^i ¼ kf À ^k2 ¼ (f À ^)T (f À ^) f f f f (6:77) The other cost function. . .7 . J2. 0 0 07 7 6. that is. Let the cost f f function J1 be a measure of closeness of f to ^. . The value of this function at n different points is given in a vector ^: f ^ ¼ ½ f ðx0 Þ f f ð x1 Þ ÁÁÁ f ðxnÀ1 ÞT (6:76) The objective is to ﬁnd a vector f that is a smooth approximation of ^. . 7 60 0 .. It is given as the distance between the second derivative of function f and zero: xnÀ1 ð J2 ¼ x0 !2 d2 f (x) dx dx2 (6:78) The cost function J2 is approximated using J2 ¼ where T Q ¼ Cn C n nÀ3 X i¼0 ðfi À 2fiþ1 þ fiþ2 Þ2 ¼ f T Qf (6:79) (6:80) and zﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄ}|ﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄﬄ{ 3 2 0 0 0 ÁÁÁ 0 0 0 0 0 0 7 6 1 À2 1 6 . 6.1 One-Dimensional Smoothing Algorithm Assume that we have a function of one variable.2 MULTIDIMENSIONAL SMOOTHING ALGORITHM The multidimensional smoothing algorithm is covered in this section.7 6 7 6 . 6... 5 4 0 1 À2 1 0 0 0 ÁÁÁ 0 0 0 0 n (6:81) .288 Control of Color Imaging Systems: Analysis and Design 6. We ﬁrst discuss the 1-D smoothing algorithm and then extend the approach to multidimensional (2-D through 4-D) case.2... .6. f(x).

Let f be the n Â m matrix that is the smooth approxif mation of ^ that we are trying to ﬁnd.153). is given as J2 ¼ nÀ2 m n mÀ2 X XÂ Ã2 X X Â Ã2 fij À 2f(iþ1)j þ f(iþ2)j þ fij À 2fi(jþ1) þ fi(jþ2) i¼1 j¼1 i¼1 j¼1 (6:86) . The vector f that minimizes J will be the best ﬁt for a given a and is obtained by setting the gradient of J with respect to f equal to zero: qJ1 ¼ 2(f À ^) f qf qJ2 ¼ 2Qf qf qJ ¼ 2(f À ^) þ 2aQf ¼ 0 f qf The solution is f f ¼ ðIn þ aQÞÀ1^ where In is an n Â n identity matrix. x and y.6. there is a function of two variables. that we are trying to approximate. the relative importance of the smoothness and goodness of ﬁt can be adjusted. which is a measure of the smoothness of f.2 Two-Dimensional Smoothing Algorithm (6:84) (6:83) In the 2-D case.Interpolation of Multidimensional Functions 289 We now deﬁne the total cost function as a linear combination of the two cost functions: J ¼ J1 þ aJ2 (6:82) where a ! 0. This function measures both the smoothness and the goodness of the ﬁt. The values of this function at evenly spaced points in the xÀy plane are stored in the n Â m matrix ^. By changing the value of the parameter a. In this case. J1 is a measure of the distance f between ^ and f and is given by f J1 ¼ n m X XÀ Á Â Ã ^ij À fij 2 ¼ k^ À f k2 ¼ Tr (^ À f )(^ À f )T f f f f F i¼1 j¼1 (6:85) where F stands for the Fibonacci norm (Equation 3.2. The second cost function J2. 6.

Let f be the f . In the 3-D case.89 into Equation 6. which can be solved by the MATLAB lyap function. we ﬁrst consider the 3-D case. that we are trying to approximate.88 results in À Á T T f aCn Cn þ In f þ f aCm Cm ¼ ^ (6:91) (6:90) (6:89) (6:88) (6:87) where In is an n Â n identity matrix. Similar to the 1-D case. Note that if either of n or m is one. The values of this function at evenly spaced points in the 3-D space are stored in an n Â m Â l tensor ^. y.6. This equation is a special case of the Sylvester equation (AX þ XB ¼ C). then this reduces to the same equation used in the 1-D case. a can be used to control the relative importance of the smoothness and the goodness of the ﬁt. J ¼ J1 þ aJ2. and z.86 is a measure of smoothness with respect to the variable x and the second term with respect to the variable y. we let the cost function be a linear combination of J1 and J2. Therefore.81 Tr stands for the trace of a matrix Since we want a ﬁt that is both smooth and approximates the data well. We then try to ﬁnd the minimum of J.86 can be written as h i h À ÁT i J2 ¼ kCn f k2 þ kCm f T k2 ¼ Tr Cn f ð Cn f ÞT þ Tr Cm f T Cm f T F F À T T Á À T Á ¼ Tr f Cn Cn f þ Tr fCm Cm f T where Cn and Cm are matrices deﬁned by Equation 6.3 Three-Dimensional Smoothing Algorithm The 4-D smoothing algorithm can easily be derived from the 3-D case.2. 6. we have a function of three variables. that is. To minimize J.290 Control of Color Imaging Systems: Analysis and Design The ﬁrst term in Equation 6. Equation 6.90 and 6. x. we set the gradient of J with respect to f equal to zero: qJ qJ1 qJ2 þa ¼0 ¼ qf qf qf since qJ1 ¼ 2(f À ^) f qf and qJ2 T T ¼ 2afCn Cn þ 2afCm Cm qf Substituting Equations 6.

J2 ¼ h f 1 Cn . This yields qJ qJ1 qJ2 þa ¼0 ¼ qf qf qf À T Á À T Á À Á f À ^ þ af 1 Cn Cn þ af 2 Cm Cm þ af 3 ClT Cl ¼ 0 f À T Á À T Á À Á f 1 aCn Cn þ f 2 aCm Cm þ f 3 aClT Cl þ Il ¼ ^ f (6:96) (6:97) (6:98) T T Since aCn Cn .Interpolation of Multidimensional Functions 291 n Â m Â l tensor that is a smooth approximation of ^. Similar to the other cases. and z. Using the discrete second derivative.99 into Equation 6. f 2 Cm i þ h f 3 Cl . i stand for tensor multiplication. aCm Cm .98 results in À Á À Á À Á ^ ¼ f 1 RuRÀ1 þ f 2 SwSÀ1 þ f 3 TcT À1 f (6:100) . and aClT Cl þ Il are all positive deﬁnite diagonalizable matrices. f 3 Cl i Similar to the 2-D case. Then. we f f deﬁne J1 as a measure of closeness of f to ^ to be J1 ¼ n m l X X XÀ Á ^ijk À fijk 2 f i¼1 j¼1 k¼1 (6:92) Let J2 be a measure of the smoothness of f. we have J2 ¼ nÀ2 m l X X XÂ i¼1 j¼1 k¼1 n m lÀ2 X X XÂ Ã2 þ fijk À 2fij(kþ1) þ fij(kþ2) i¼1 j¼1 k¼1 fijk À 2f(iþ1)jk þ f(iþ2)jk Ã2 þ n mÀ2 l X X XÂ i¼1 j¼1 k¼1 fijk À 2fi(jþ1)k þ fi(jþ2)k Ã2 (6:93) The three terms in the above equation are measures of smoothness with respect to the three variables x. respectively. Let n stand for the n-mode multiplication of a tensor with a matrix and h . The deﬁnitions of these operations are given in Appendix C. we can deﬁne T aCn Cn ¼ RuRÀ1 T aCm Cm ¼ SwSÀ1 aClT Cl þ Il ¼ TcT À1 (6:99) where u. let the overall cost function be J ¼ J1 þ aJ2 (6:95) (6:94) The minimum of J is obtained by setting the gradient of J with respect to f equal to zero. f 1 Cn i þ h f 2 Cm . w. y. and c are diagonal matrices. Note that it is an approximation of the integral of the square of the second derivative of f along each of the three directions. Substituting Equation 6.

T and aCl Cl þ Il . The algorithm for 4-D is as follows: T T T Step 1: Form and diagonalize the positive deﬁnite matrices aCn Cn . aCk Ck . aCm Cm . .292 Control of Color Imaging Systems: Analysis and Design f Multiply both sides of Equation 6. ^ F ¼ F 1 u þ F 2 w þ F 3 c Therefore.100 by the operator 1 RÀ1 2 SÀ1 3 TÀ1 ^1 À Á À Á RÀ1 2 SÀ1 3 T À1 ¼ f 1 uRÀ1 2 SÀ1 3 T À1 þ f 2 wSÀ1 1 RÀ1 3 T À1 À Á þ f 3 cT À1 1 RÀ1 2 SÀ1 (6:101) Deﬁne ^ f F ¼ ^ 1 RÀ1 2 SÀ1 3 T À1 and F ¼ f 1 RÀ1 2 SÀ1 3 T À1 Then. where Il is the l Â l identity matrix: T aCn Cn ¼ RuRÀ1 T aCk Ck ¼ TcT À1 T aCm Cm ¼ SwSÀ1 aClT Cl þ Il ¼ VgV À1 (6:108) Step 2: Compute ^ f F ¼ ^ 1 RÀ1 2 SÀ1 3 T À1 4 V À1 Step 3: Compute Fijkl ¼ ^ Fijkl uii þ wjj þ ckk þ gll (6:110) (6:109) Step 4: Find f using Equation 6. ^ Fijk ¼ Fijk uii þ Fijk wjj þ Fijk ckk Fijk ¼ ^ Fijk uii þ wjj þ ckk (6:105) (6:106) (6:104) (6:103) (6:102) Once F is determined.107. f can be calculated from f ¼ F 1 R 2 S 3 T (6:107) We now extend the theory to the 4-D case.

consider the image reproduction system shown in Figure 6. the proﬁling LUT is a mapping from the 3-D device-dependent sRGB to the 4-D device-dependent CMYK color space.22 Image reproduction with LUTs.26.6. mean. TABLE 6.005 2.17 Simulation Results of 3-D Proﬁling Mean DE2000 No noise LUT Noise and no smoothing Noise and smoothing 0. the input sRGB image is mapped to the CMYK color space by a multidimensional proﬁle LUT. The lower the DE2000. 6.22.18 correspond to the optimal choice of a ¼ 0. To produce a smooth characterization LUT.17. The results of simulations of the Lena image thorough the reproduction system of Figure 6. and 95th percentile DE2000 for different values of a are shown in Figure 6. In this example. there is a 50% improvement in the 95th percentile. DE2000 is used as a measure of the performance. the ﬁnal LUT may not be a smooth 3-D function. The smoothing parameter a is changed from 0 to 2.Interpolation of Multidimensional Functions Calibrated C'M'Y'K' L*a*b* Printer 293 sRGB CMYK With smoothing algorithm CMYK Calibration sRGB ΔE2000 computation Calibrated sRGB CMYK CMYK C'M'Y'K' Calibration Printer Without smoothing algorithm L*a*b* ΔE2000 FIGURE 6.891 1. As can be seen. It is computed by converting the input sRGB image to L*a*b* and comparing the L*a*b* to the printer output L*a*b*.3 APPLICATION TO PRINTING SYSTEMS As an example. Since color measurement is a noisy process. The original and reproduced images with and without smoothing are shown in Figures 6.22 with and without a smoothing algorithm for a noisy sRGB to CMYK LUT are shown in Table 6.23 through 6. we insert a 3-D smoothing algorithm.25.419 95th Percentile 1.609 Maximum DE2000 6. In this case.999 7. the better the color balance and image quality become. The results of the DE2000 accuracy in terms of maximum. The results shown in Table 6.954 12. which is a GCR-constrained inverse of the printer (Chapter 7). The CMYK pixels can be processed through 1-D calibration tone reproduction curves to generate the printer C0 M0 Y0 K0 .664 6.5.465 2.923 .

18 One-Dimensional LUT x y 0 2.3 1.6 9 0 Original image Reproduced image FIGURE 6.5 2. Original and reproduced image with .294 Control of Color Imaging Systems: Analysis and Design TABLE 6.5 3.2 3 4.) noisy proﬁle LUT.9 7 0.5 2 6.24 (See color insert following page 428.23 (See color insert following page 428. Original and reproduced image with Original image Reproduced image FIGURE 6.) noise-free proﬁle LUT.9 8 0.2 5.6 4.5 4.8 6 1.

5 2 2.26 Maximum. Original and reproduced image with 14 12 10 8 6 4 2 0 Mean ΔE 2000 Max ΔE 2000 95th percentile ΔE 2000 0 0.5 1 1.5 α 3 3.5 4 4. mean.Interpolation of Multidimensional Functions Original image Reproduced image 295 FIGURE 6.5 5 FIGURE 6.25 (See color insert following page 428. and 95th percentile DE2000 for different values of a.1 Consider the 1-D function f (x) ¼ ln (x þ 1) 0 x 1 .) noisy proﬁle LUT and smoothing algorithm. PROBLEMS 6.

What is the resulting MSE? 6. What is the resulting MSE? 6. y.75. The total number of grid points is 100. What is the resulting MSE? (b) Repeat part a using cubic interpolation. The total number of grid points is 1728. (a) Use linear interpolation to estimate the value of the function at x ¼ 3.75. (a) Use trilinear interpolation to ﬁnd the L*a*b* values corresponding to the following CMY patches: Patch 1: Patch 2: Patch 3: Patch 4: CMY1 ¼ [20 20 CMY2 ¼ [40 40 CMY3 ¼ [60 60 CMY4 ¼ [80 80 20] 40] 60] 80] (b) Is this printer gray balanced? .2 Consider the 1-D LUT shown in Table 6.5 Consider a CMY printer characterized through the 3-D LUT given in Table 6.3 Consider the 2-D function given by À Á z ¼ f (x.568.68.19.896. y) ¼ x þ y þ xy1:8 ln (x þ 1) 0 x.18. What is the resulting MSE? (c) Repeat part b for the point with coordinates x ¼ 0. (a) Use linear interpolation to estimate the values of the function at x1 ¼ 0. y ¼ 0.25 and y ¼ 0. and z ¼ 1. y ¼ 0. y 1 (a) Create a 2-D LUT of size 10 Â 10 by uniformly sampling the function from 0 to 1 in both directions. 6. y. z 1 (a) Create a 3-D LUT of size 12 Â 12 Â 12 by uniformly sampling the function from 0 to 1 in all directions. z) ¼ (xz þ yz þ xyz þ 1) ln (xyz þ 1) 0 x. (b) Use trilinear interpolation to estimate the values of the function at x ¼ 0. (b) Repeat part a using (i) The Shepard technique (use L1 norm and m ¼ 1) (ii) The moving matrix with parameters m ¼ 2 and e ¼ 0 6.296 Control of Color Imaging Systems: Analysis and Design Create a 1-D LUT of size 10 Â 2 by uniformly sampling the function from 0 to 1.68. (b) Use bilinear interpolation to estimate the values of the function at x ¼ 0.75.8.234 and x2 ¼ 0. and z ¼ 0.4 Consider the 3-D function g ¼ f (x.

37 63.00 À7.72 44.64 À20.40 57.10 68.75 27.47 À44.59 7.42 À1.65 29.25 13.95 À6.88 À7.19 Three-Dimensional LUT C 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 85 85 85 85 85 85 85 85 85 85 85 85 85 85 85 85 170 170 170 170 170 170 170 170 170 170 M 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 0 0 0 0 85 85 85 85 170 170 Y 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 L* 100.02 À13.42 5.96 À39.91 b* 0.66 68.53 37.61 96.80 53.14 57.25 9.03 45.61 À26.37 61.69 54.30 65.82 17.73 23.28 24.24 À18.72 19.00 62.18 78.56 À16.84 78.74 À17.93 À29.Interpolation of Multidimensional Functions 297 TABLE 6.76 16.94 80.47 43.90 68.92 49.89 42.10 11.60 31.00 35.31 40.02 32.96 82.65 56.03 À20.21 41.18 87.86 68.85 À14.95 À11.70 81.53 45.02 94.64 23.35 47.50 67.79 48.84 44.96 53.16 83.44 À11.16 86.26 À30.82 66.01 54.86 23.11 64.82 54.76 44.07 55.98 À9.95 18.19 À46.73 67.58 55.83 À0.33 41.82 À41.70 53.8 (continued) .49 80.51 69.81 2.54 77.67 a* À0.35 À28.85 69.53 25.72 100.65 1.42 13.12 30.92 54.67 57.44 12.99 À14.25 73.98 81.00 97.59 29.53 À8.31 À45.07 15.72 56.16 70.49 49.04 À4.90 À43.21 69.69 À24.13 79.65 54.60 66.47 70.41 À25.95 55.45 45.

46 0.57 13.22 À57.68 23.48 37.8 Assume that we have a CMY ! L*a*b* printer characterization function with Jacobian J at the nominal value CMY ¼ [120 67 145] given by 2 À0:1586 À0:1223 qLab J¼ ¼ 4 À0:1267 0:3146 qCMY À0:1292 À0:0480 3 À0:0075 À0:0299 5 0:4026 .40 39.89 À47.78 À34.30 1.10 44.20 53.37 23.14 À22.54 44.6 Repeat Problem 6.91 À3.25 À57.77 À55.06 6.72 À56.10 35.79 54.71 À10.77 40.50 À55.76 54.66 À38.39 45.20 34.44 26.15 À16.49 54.41 45.20 38.12 42.13 À24.92 11.7 Show that the Shepard interpolation algorithm given by 8 PN Àm < i¼1 yi di PN Àm ^¼ y d : i¼1 i yj if di 6¼ 0 if dj ¼ 0 for all i for some j will result in a differentiable function if parameter m > 1.42 À5.02 6.56 54.28 35.77 14.04 a* 9.07 34.12 24.51 À30.55 b* 5.17 33.40 À30.01 27.51 À27.94 46.99 28.3 using (a) Tetrahedral interpolation (b) Shepard interpolation (c) Moving matrix 6.18 À51.298 Control of Color Imaging Systems: Analysis and Design TABLE 6.21 À2.48 À7.64 36.50 À24.80 32.98 26.67 37.67 À35.60 45.79 À61.72 À12.80 30.19 (continued) Three-Dimensional LUT C 170 170 170 170 170 170 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 M 170 170 255 255 255 255 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 Y 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 L* 45. 6.05 4.

74 0 14.20.Interpolation of Multidimensional Functions 299 Find the range of parameter m that guarantees convergence of the ICI algorithm.39 35.68 87.27 À68.82 À7.59 32.76 60.87 36.2 18.25 23.53 À88.37 À50. (a) Using trilinear interpolation.91 À9.18 À26.11 46.76 64.52 À52.36 26.86 24.85 À84.99 89.41 À17.35 À6.13 39.28 48.52 À64.22 30.83 61.69 43. ﬁnd the L*a*b* values corresponding to the following RGB values: RGB1 ¼ [23 23 23] RGB2 ¼ [123 123 50] TABLE 6.69 À27.23 34.43 87.03 À72.81 32.53 À33.9 Consider a scanner characterization color space transformation from a calibrated scanner RGB to a device-independent color space L*a*b* given by the 4-level (64 grid points) LUT shown in Table 6.76 (continued) .49 0 À45.52 61.59 67.47 62.9 19.68 b* 0 À48.73 À22.38 À10.17 83.41 À51.09 63 66.93 À41.7 À108 37.22 40.86 À37.88 À38.87 À55.17 À79.05 15.15 40.92 À5.79 0.06 91.2 42.69 37.33 58.51 53.21 62.43 À101. 6.12 36.32 33.25 36.42 À50.58 26.09 64.9 a* 0 23.98 À49.56 À75.71 À14.28 63.20 RGB to L*a*b*LUT R 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 85 85 85 85 85 85 85 85 85 85 85 85 G 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 0 0 0 0 85 85 85 85 170 170 170 170 B 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 L* 0 5.61 32.

34 40.06 À44.53 À7.51 67.66 89.56 91.99 69.1 À12.03 36.13 72.88 80.31 À30.20 (continued) RGB to L*a*b*LUT R 85 85 85 85 170 170 170 170 170 170 170 170 170 170 170 170 170 170 170 170 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 G 255 255 255 255 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 0 0 0 0 85 85 85 85 170 170 170 170 255 255 255 255 B 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 0 85 170 255 L* 88.33 65.96 À18.61 72.76 46.97 91.59 À60.23 67.17 0 11.8 83.71 0 (b) Using the moving-matrix technique.55 39.23 À64.39 0 À42.32 67.02 32.300 Control of Color Imaging Systems: Analysis and Design TABLE 6.67 75.38 62.46 85.51 77.4 46. 6.13 69. ﬁnd the RGB values corresponding to L*a*b* ¼ [60 82 À 62].08 79.64 49.98 92.81 76.03 À26.99 16.89 À37.43 À7.41 0 b* 84. estimate the Jacobian of the transformation at the nominal value RGB ¼ [120 67 145] using .78 À44.78 À11.69 91.92 89.39 60.61 65.14 28.85 49.46 À44.43 94.96 À29.48 40.82 55.56 69.12 87.47 54.35 97.7 61.1 56.77 82.76 À9.97 94.55 54.66 70.54 77.75 43.07 27.11 62.14 76.61 66.15 97.22 36.31 100 a* À67.2 36.97 70.61 88.85 28.42 98.42 59.39 À9.82 53.13 42.63 À52.66 À8.23 65.31 66.79 À12.92 35.41 32.4 54.8 60.7 45.51 28.53 52.45 12.77 À83.9.43 À71.10 (a) Using the RGB to L*a*b* LUT of Problem 6.62 2.46 34.93 À57.69 À42.

5. Koditschek.752. 499–504. M. REFERENCES 1. The use of spectral regression in modeling halftone color printers. Proceedings of the ACM National Conference. G. Journal of Optical Engineering. US Patent 7. 10. 18–23. US Patent 7.K.K. The Jacobian is deﬁned by 2 3 qL*(k) qL*(k) qL*(k) qR(k) qG(k) qB(k) qL*a*b*(k) 6 qa*(k) qa*(k) qa*(k) 7 ¼ 6 qR(k) qG(k) qB(k) 7 Jk ¼ 4 5 qRGB(k) qb*(k) qb*(k) qb*(k) qR(k) qG(k) qB(k) (b) Find the range of parameter m that guarantees convergence of the ICI algorithm that can be used to ﬁnd the inverse LUT. A two-dimensional interpolation function for irregularly-spaced data. pp. Portland. Kelvin E. Jul. Bala. Gray balance control loop for digital color printing systems. Matthew. and A. Mestha.E. March 26–29. 2000. 3. S. Baltimore. 1996. .K. IEEE International Conference on Neural Networks. SPIE Press. MD. Dynamic optimization algorithm for generating inverse printer map with reduced measurements. 2003. March 2003. Bala. Rochester. Speech and Signal Processing. May 14–19. Mestha.N. Optic=Imaging in the Information Age. John Wiley & Sons Inc. PICS 2000 Conference. R. 2006 IEEE International Conference on Acoustics. Proceedings of SPIE. MD. Chong and S. FL. P. J.581. Xia. Tekalp. Y. 2000. New York. 700–716. Piecewise linear homeomorphisms: The scalar case.Z. 2005. 334–340. 2-D and 3-D printer calibration algorithms with printer drift. 2658. Bellingham. CRC Press.A. and S. 18–23. Dianat. pp. 3. Apr. 1968. 16. G.R. Jul. 9.. Khargonekar. S. D. Baltimore. 372–375.S. Kang.A.K. IEEE Transactions on Image Processing. and M. Boca Raton. Chang. 2005. Shepard. R. Comparison of 1-D. M. 2008. Groff. 259–264..H. Toulouse.K. 12. Reﬁnement of printer transformation using weighted regression. 11. Sequential linear interpolation of multidimensional functions. Dianat. End-to-end color printer calibration by total least squares regression. 13. R. 2008. NIP21: Proceedings of the IS&T’s International Conference on Digital Printing Technology. 4. 1996.397.. E. Viturro.355. L. and L. pp. 2006. IEEE Transactions on Image Processing. OR. Oct. 5. New York. D. 1999. WA. R. IST=OSA Annual Conference. pp. Color Technology for Electronic Imaging Devices.K. E. 15.Interpolation of Multidimensional Functions 301 numerical differentiation and trilinear interpolation. L. Viassolo and L. Sept. Two-dimensional calibration architectures for color devices. 517–524. R. 8. 8. 1996. and R. 505–510. 1997.R. 625. 6. An Introduction to Optimization.G. Zak. Balasubramanian and M. Wang. Maltz. France. L.E. A practical algorithm for the inversion of an experimental input-output color map for color correction. Balasubramanian..R. NIP21: Proceedings of the IS&T’s International Conference on Digital Printing Technology. Van de Capelle. Mestha.K.S. Sharma. Sept. 14. 8. Gurram. Sharma. Geoffrey J. Malz. 6.E. Dianat. and Edward J. L.K.P. J. 2. Mestha. Dianat. Mestha and S. TRC smoothing algorithm to improve image contours in 1-D color controls. G.W. Sharma. Sept. 1996.A. and P. Reference input=output medium metric RGB color encodings (RIM=ROMM RGB). Saber. 7. 1231–1245. D. H. Mestha. Digital Color Imaging Handbook. 42(3).

.

g. Numerous techniques for device characterization.1 INTRODUCTION A color management architecture for a digital imaging system provides a means for processing digital images such that the colors produced on the output devices are a reasonable representation of the colors in the input image ﬁle. display devices (e. For each device class.g. CDs.. and output devices (e.. A key output from this chapter is the well-tuned custom multidimensional lookup table (LUT). display. The preferred industry standard used by many production print shops for managing color is based on standards set by the ICC. RGB monitors and RGB projectors). These tables can be used to color-manage input images so that when the resulting images are printed on output devices.7 Three-Dimensional Control of Color Management Systems 7. Some of these techniques use advanced control-based approaches. and rendering devices [1].g.2 IMAGE PATH ARCHITECTURE The control of color is a system-wide problem. RGB scanners. The algorithms described in this chapter use theoretical tools from previous chapters to produce multidimensional color transforms for managing color between devices. printers). and constrained inverse and gamut mapping are described in detail. The International Color Consortium (ICC) was set up to provide a standard paradigm for managing color with image capture.. The ICC speciﬁcation divides color devices into three broad categories: input devices (e. 7. RGB cameras. An image ﬁle may be created using a computer or may be captured by other imaging devices such as a scanner or a digital camera. better in terms of accuracy and consistency as compared to prints from nonoptimal color management systems. gray component replacement (GCR). The color strategy followed by print shops must use standards that provide a link between color transforms and numerous devices so that the colors represented in different device-speciﬁc color spaces can be transformed to a space local to the rendering device. electronic documents). a 303 . they exhibit increased color quality. The ICC speciﬁcation is a vendor-neutral cross-platform color management standard created by a group of industry color experts to improve color workﬂow and is implemented in an ICC proﬁle format.

These L*a*b* or XYZ values are passed to the destination proﬁle. This situation is called the gamut-mapping problem and it is normally handled inside the destination proﬁle. CMYK) to another output device color space (CMYK) without going through the PCS.304 Control of Color Imaging Systems: Analysis and Design series of base transforms are included in the ICC proﬁle that perform the conversion between different color spaces. Although the destination proﬁle is simply an XYZ or L*a*b* (input of Figure 7. which deﬁnes yet another mathematical transformation (e. A color lookup table (CLUT) or a 3DLUT is a major component of the destination ICC proﬁle that is a GCR-constrained. measurement geometry.8 for detailed steps involved in the creation of a colorimetrically accurate destination proﬁle. perceptual. and destination architecture. a three-to-three (RGB-to-L*a*b*) or a four-to-three transform (CMYK-to. and blue light seen by the source camera) and a reference color space. In general. in this chapter we focus our algorithms on the generation of an accurate colorimetric proﬁle.. the basis for creating a device link proﬁle or proﬁles with different rendering intents has always been an ordinary ICC proﬁle such as the colorimetric proﬁle. the proﬁle is divided into several components as shown in Figure 7.1) to CMYK transform (output of Figure 7.g. in particular. describes the color as perceived by the human eye. This type of division will help the process of creating a destination proﬁle. a tone reproduction curve (TRC).. etc.1:2004-10 speciﬁcations because of the need for tuning the proﬁles to speciﬁc applications. gamut-mapped LUT from RGB triplets to device CMYK space (Figure 7.2). In the ICC proﬁle architecture a reference color space in L*a*b* or XYZ called the proﬁle connection space (PCS) [1]. green. Hence. Also. and a second TRC. a three-to-four transform used for a CMYK printer) between the amounts of the colorants from reference color space and the destination color space that the device would need to use to render the color. is assigned to each image ﬁle that deﬁnes a mathematical transform between the numbers in the input ﬁle (i. A source proﬁle. Device Link proﬁles are another special kind of ICC proﬁles that provide a dedicated transformation from one input device color space (e.) included in the ICC. these mathematical transformations are embodied as multidimensional LUTs.L*a*b*). See Section 7. Thus the ICC proﬁle format is made generic to the extent that each application vendor can foster their own proprietary technology to differentiate their art from their competitors. a 3DLUT. the amount of red. a destination proﬁle. and illuminants. speciﬁed in XYZ or L*a*b*. reference. However.e. . Using this kind of source. The destination proﬁles have an additional complexity because they must also handle requests for colors the destination device cannot make. the source and destination devices from different vendors can be used in a system to produce the color with good results that is most suitable for a particular application. in order to obtain the biggest possible color gamut with minimum amount of toner and improved color rendition accuracy. This reference color space.1). is used for standard observers.. Figure 7. there can be many different rendering intents (colorimetric.1 shows a footprint of the destination ICC proﬁle used with various components to translate color information from the PCS to the device-speciﬁc CMYK space.g. saturation.1: the XYZ or L*a*b* to RGB matrix.

etc. etc. (ii) system errors like halftone noise. ProPhoto RGB Gamut mapping Lab GCR constrained LUT (RGB->Lab -> CMYK) CMYK TRC C M Y K FIGURE 7.’’ In technical terms. That means the desired perceptual color difference for in-gamut colors between the input L*a*b* values to the rendered * L*a*b* values of the pixels must be less than or equal to 1 (DEab. The simplest requirements in device owners terms are ‘‘get the color right. capturing variation in temperature. (iii) sensor .3 PROFILING—A COMPLEX SYSTEM PROBLEM While multidimensional proﬁles offer opportunity to improve color quality of output devices. This is extremely difﬁcult to achieve due to (i) 8 bit quantization limits of the imaging system.. However. DE2000 .1 LUT B2A table converting PCS to device CMYK space in the ICC proﬁle. and all the time. media. the complexity involved in the creation of these proﬁles is high. Colors have to be within perceptual limits in all regions of the color space that the printer can produce. everywhere.. RGB e. As seen in the discussion below. humidity.1 TIGHT COLOR RENDITION REQUIREMENTS Very accurate colorimetrically matched reproduction is desired for high quality rendering.) unit. 7.2 LUT B2A table converting PCS to device CMYK space in the ICC proﬁle (Note: RGB is an intermediate color space). In general. 7. the phrase ‘‘every time.g. ‘‘getting the color right’’ on multiple devices is infeasible because different devices have different color gamuts and it is physically impossible for a device to produce a color outside its reproducible gamut. the requirements for high quality rendering are very rigid. and the best proﬁles. ProPhoto RGB] FIGURE 7. even with a stable device.g.. it is not easy to ‘‘get the color right’’ so that the color differences are maintained within the perceptual limits. development noise.3. everywhere and all the time’’ implies achieving device process stability under all the operating environments.Three-Dimensional Control of Color Management Systems This space is selected to maximize the gamut coverage RGB TRC 3DLUT CMYK TRC C M Y K 305 Profile connection space (PCS) RGB X or L* 3×3 Y or a* Matrix [e. every time.

For instance. Each multidimensional transform must utilize various methods to achieve the desired color and also give experienced users the opportunity to control the gamut-mapping behavior. and this requires rules to choose from the many possible CMYK formulations (called gray component replacement [GCR]) and inversion methods. The paper blue may not be as saturated as the one displayed on the monitor. printing a mostly saturated blue color.3 of a printer. the proﬁling algorithm errors can make it impossible to achieve required rendition accuracy since the proﬁle algorithms use a many-to-one inversion. In spite of improvements with respect to (i). some colors outside a destination device gamut need to be mapped to the inside of the gamut or on the boundary as they otherwise cannot be reproduced on the output device and would simply be clipped.3. one must ﬁnd the right CMYK formulation for each color in the table. these image quality defects are related to gamut mapping in particular regions of the color space (e. Otherwise. Contours can occur due to a many-to-one mapping at the intersection between mapped colors and in-gamut colors near the gamut boundary. out of gamut colors are mapped to the destination device gamut that they are able to produce. on paper using a typical CMYK printer will likely fail. dark brown region). output images can have contours.306 Control of Color Imaging Systems: Analysis and Design errors.3 SMOOTHNESS Reproduction of color sweeps that cross the gamut boundary must be smooth. (ii). 7. Vector plot of out-of-gamut colors shown with respect to reproducible gamut . and (iv) proﬁling algorithm errors. Conversely..3. the bright cyan of an inkjet printer may not be easily presented on an average computer monitor. 7. Since many devices do not have full gamut coverage.2 GAMUT LIMITATION A device must be able to produce good color from input image ﬁles that contain colors outside the nominal device gamut (Figure 7.g. as displayed on a monitor. Generally.3). To build a multidimensional LUT. FIGURE 7. and (iii). For example.

then the colors initially matched to the original print are no longer valid. leading to unpredictable results.4 ICC WORKFLOW The multidimensional proﬁle LUTs embedded in ICC proﬁle have ﬁnite nodes at which device values are available. halftone. All of these steps are too demanding for high-end printing since even a small numerical error can lead to a noticeable change in output color and a degradation in the image quality. linear interpolation provided by ICC image path may not provide sufﬁcient accuracy.5 ENGINE CONDITIONS A stable engine condition is essential to repeatedly produce the color in the original input ﬁle for the same media.Three-Dimensional Control of Color Management Systems 307 7. The ﬁrst step in the proﬁle creation stage is device characterization. static gamut mapping is no longer optimal.4 Nonlinear color surfaces (CYK L*a*b* gamut with color surfaces for CY planes for ﬁxed K and M ¼ 0). Nonlinear color surfaces FIGURE 7.4. As gamut boundaries move with time. Nowadays. even custom proﬁles with static gamut mapping will not reproduce optimal output quality. if the overall gamut of the engine varies. Also. 7. toner. The ICC path provides simple linear interpolation with limited resolution for processing pixels at high speed.3. Therefore. When device behavior is nonlinear as for the gamut of Figure 7. We show the adaptation process using recursive least-squares (RLS) identiﬁcation algorithms. Input image ﬁles with color pixels not on the nodes will require interpolation. in the following section we ﬁrst describe the underlying theory and algorithmic details of various methods applicable to characterizing a four-color print engine. The next step involves computing the GCR-constrained inverse and gamut mapping the out-of-gamut colors to the characterized boundary. .3. characterization is performed by printing and measuring a set of color targets. etc.

1 LEAST-SQUARES ESTIMATION We will consider two parts to the empirical or interpolation-based approach to printer modeling. Using the historical data and the new measurements. and the predicted colors from the adapted model can better represent the real system output at that time. a more elaborate. Once an optimal parameterized model is selected. The adaptation process [11] is implemented using measurements from the sensor to adjust or tune the model parameters over time. It accounts for variations in printer state over time t0 þ t. an adaptive algorithm will estimate the model parameters more accurately.4. or parameterized analytical functions that ﬁt the data. which is typically done (at the printer manufacturer) ‘‘in the factory’’ but could also be performed ‘‘in the ﬁeld’’ if an in-line color device is available [7–10]. Both kinds of approaches are capable of predicting the color response of the device for a variety of input images. errors in capturing the actual physical process in the presence of device drift over time. although we cannot use an empirical model to explain a system. A mathematical model is generated and initial parameters are selected from this data using data ﬁtting techniques such as the least-squares method. Accurate ﬁrst principle models (so-called ‘‘white box’’ models) are not available for all kinds of imaging devices. Data is used to assess the structure of the functional form and then estimate its parameters. the second part involves an adaptation process. . This type of model can help us to inject meaningful time-varying effects into the system. which essentially treat the device as a black box with inputs and outputs and (2) analytical or ﬁrst principle approaches which attempt to characterize the device color response using the fewest number of measurements to arrive at analytical functions that have a physical connection to the process. this kind of empirical modeling of a dynamic system is called system identiﬁcation. In this section. light scattering effects. The ﬁrst part is to obtain the model at time t0. we present several empirical models with approaches to capture both time-zero and time-varying effects.4 CHARACTERIZATION OF COLOR SYSTEMS Conventionally. parameterized nonlinear spectral model of the printing system incorporating reasonable abstractions of the process is described. Thus. These models may contain only nonparametric LUTs. The empirical or interpolation-based methods are generally measurement intensive and require the use of a large set of experimentally generated data between inputs and outputs [2–6]. The complexity of models. and many other uncertainties associated with the physical device itself make it impossible to perfectly model the device over a reasonable period of time. The function passes through the data points approximately. One can understand that measurement data is crucial for an empirical model. there are two different approaches to color modeling: (1) empirical or interpolation-based approaches.308 Control of Color Imaging Systems: Analysis and Design 7. In Chapter 10. This approach may require measurements of anywhere from one to thousands of color patches. 7. we can use such a model to predict the behavior where data does not exist. In automatic control literature.

quadratic.g. etc. therefore. noninvertible. linear. A ¼ A0 ¼ ½1 u1 u2 u3 u4 . the parameters contained in u0 cannot be solved using simple linear algebra or matrix methods. the sample output test color data is represented in L*a*b* and the corresponding input value is represented in CMYK. For example. We can then write parameters of the linear afﬁne model (Equation 7. For a scalar case. If there are N colors in the training set.g. or cubic terms is introduced below. afﬁne. then an input–output model can be written as y ¼ u1 f1 (u) þ u2 f2 (u) þ Á Á Á þ un fn (u) where y is the output (sensor) data u is the input (e. y ¼ Au (7:2) (7:1) where the matrix A is named the regression matrix and it contains the form of the model and u is the matrix containing the parameters of the model.. linear. and u is the input.) In vector and matrix form.2) as follows: y ¼ y0 ¼ ½y1 y2 2 y3 . and y0 ¼ A0 u0 7 7 M23 M33 5 M24 M34 b1 6 6 M11 6 u ¼ u0 ¼ 6 M12 6 6 4 M13 M14 where u1 ¼ c u2 ¼ M u3 ¼ Y u4 ¼ K y1 ¼ L* y2 ¼ a* y3 ¼ b* bi . Mij are constants (7:3) Note. CMYK values for a printer) data ui is the parameter that needs ﬁtting or identiﬁcation fi (u) is the known form of the model (e. Since the A0 matrix is nonsquare and. quadratic. then y is a matrix of size N Â 3.1. . if y is the output data from the sensor.Three-Dimensional Control of Color Management Systems 309 7. A0 is a matrix with N number of rows.4.1 A Linear in the Parameters Model A simple ‘‘linear in the parameters’’ model with afﬁne. when modeling a CMYK to L*a*b* input–output data. For a linear afﬁne model (as shown) A0 will be of size N Â 5. 3 b2 b3 7 M21 M31 7 7 M22 M32 7. we used the subscript ‘‘0’’ to distinguish the variables between initial data and the updates coming at future time during the adaptation process.. First we form a residue equation.

1. It can be used as a onetime estimation algorithm to improve the numerical accuracy of a least-squares model. u4 ¼ K. the partial quadratic model can be modeled as Â Ã (7:7) A0 ¼ 1 u1 u2 u3 u4 u2 u2 u2 u2 1 2 3 4 This is yet another form suitable for linear regression. Most adaptive algorithms are of the following form: xkþ1 ¼ xk þ ae Color sensing device (7:8) Printing device CMYK values for target color Parameterized model – + FIGURE 7.5 for a test color. This process yields the standard least-squares solution. For example. À ÁÀ1 u0 ¼ AT A0 AT y0 0 0 À ÁÀ1 (7:6) The term AT A0 AT is called the pseudo-inverse of matrix A0. . other parameter models can also be built easily.7) when N number of test colors are used. u2 ¼ M.6 is used 0 0 to obtain an initial guess of the parameter matrix using the measured L*a*b* values for test colors whose CMYK values are included in the A0 matrix as u1 ¼ C. u3 ¼ Y.5 with respect to u0 and set the resulting equation to zero and solve for u0 . Once the sensor data become available. Equation 7. S ¼ rr T ¼ ðy0 À A0 u0 Þðy0 À A0 u0 ÞT (7:4) (7:5) Now minimize S by differentiating Equation 7.2 Recursive Least-Squares Estimation Algorithm The RLS algorithm is an adaptive learning method that can be used to reﬁne the parameters of the model and perform system identiﬁcation with the data collected with in situ sensors for a time-varying print engine. Using this technique. it allows the system to update the parameters to reduce the error between the outputs of the linear model and the actual printer to an acceptable level. 7. A0 is a matrix of size N Â 9 (as shown in Equation 7.4.310 Control of Color Imaging Systems: Analysis and Design r ¼ y 0 À A 0 u0 and then form the sum of the squares of the residues. Note. A schematic of the adaptation process is shown in Figure 7.5 System block diagram showing adaptation process used to tune the parameters of the linear model.

are built by using all the available input and output data points as in Equation 7.2) to obtain the overdetermined linear equation.11.10. y¼ a y (7:9) (7:10) The updated estimates. we have À ÁÀ1 Á À ÁÀ1 À T A0 y0 þ a y ukþ1 ¼ AT A AT y ¼ AT A0 þ aaT 0 (7:14) Matrix inversion lemma (Equation 5. Rewriting Equation 7. In the adaptation context.6.2. Few intermediate steps are necessary before the new estimates are determined. y ¼ Au This equation is similar in structure to Equation 7.64) states that for any matrices A.11. the equation for the new estimates again has similar structure as the initial estimate Equation 7. A more efﬁcient way of calculating the new solution utilizes Equation 7.6. B. The previously calculated inverse from Equation 7.11. let aT represent the regression matrix for the next set of data while is the corresponding output. we will need to calculate the inverse again as new data arrive.11. we refer to uk as an old estimate.6 is in fact used in the new Equation 7. This means that unnecessary computation is required to solve Equation 7. However. " !T ! #À1 ! ! À T ÁÀ1 T A0 A 0 y0 A0 (7:11) ukþ1 ¼ A A A y ¼ aT aT aT y As expected. we can write y AT A ¼ AT A0 þ aaT . When the new data arrive. and uk þ 1 as the new estimate. uk þ 1. This additional data can be incorporated using the y original equation (Equation 7.Three-Dimensional Control of Color Management Systems 311 where xkþ1 ¼ new estimate xk ¼ old estimate a ¼ weight factor ! ! measured model e¼ À values output The initial (estimated) parameters in the u0 matrix are obtained from Equation 7.6. with ! ! y0 A0 A¼ T . From simple matrix manipulation of Equation 7. and C of appropriate size À ÁÀ1 (A þ BC)À1 ¼ AÀ1 À AÀ1 B I þ CAÀ1 B CAÀ1 (7:15) . which is shown next. AT y ¼ AT y0 þ a 0 0 À T ÁÀ1 À T ÁÀ1 and P0 ¼ A0 A0 Also let P ¼ A A (7:12) (7:13) Further manipulation is done as follows to reach a simple form for the new u matrix.

The matrix P0 is updated in Equation 7. u0 ¼ P0 AT y0 and from Equation 7.312 Control of Color Imaging Systems: Analysis and Design Let AT A0 ¼ A. we get Á Â Ã À y ukþ1 ¼ I À KaT P0 AT y0 þ a 0 y y ¼ P0 AT y0 þ P0 a À KaT P0 AT y0 À KaT P0 a 0 0 (7:19) But.20 is replaced with y in Equation 7. To show the iteration steps. P0 a ¼ K[1 þ aTP0 a].16 becomes À AT A0 þ aaT 0 ÁÀ1 Â Ã P ¼ I À KaT P0 (7:18) Substituting Equation 7.17.21 for each iteration. 0 Substituting these relationships in Equation 7.8.20) is very similar to the general adaptive algorithm of Equation 7. À Á (7:22) ukþ1 ¼ uk þ Kkþ1 ykþ1 À aT uk kþ1 Kkþ1 ¼ Pk akþ1 1 þ aT Pk akþ1 kþ1 and Â Ã Pkþ1 ¼ I À Kkþ1 aT Pk kþ1 (7:23) Note: in Equation 7. it can be shown Kkþ1 ¼ Pkþ1 akþ1 (7:24) .18 into Equation 7.19. let us replace the subscripts ‘‘0’’ with ‘‘K’’ to represent kth update and use ‘‘K þ 1’’ to denote the new updates. B ¼ a. again.21 then yield the following form for the old estimate and the new estimate. Then substituting into Equation 7. we removed kþ1 from the notation. (Note: for convenience. Equations 7. Simplifying Equation 7.15.14. recognizing aT P0 a is a scalar and then substituting K Equation 7. from Equation 7. the weight matrix K and the matrix P are given by K¼ P0 a 1 þ aT P0 a and Â Ã P ¼ I À KaT P0 (7:21) (7:20) Notice that the new equation for u (Equation 7.) À Á y u ¼ u0 þ K À aT u0 where. and C ¼ aT.20 and 7.6. we get 0 À T ÁÀ1 À ÁÀ1 À ÁÀ1 À ÁÀ1 À1 T À T ÁÀ1 A0 A0 þ aaT ¼ AT A0 À AT A0 a I þ aT AT A0 a a A0 A0 (7:16) 0 0 0 Let us deﬁne K¼ P0 a 1 þ aT P 0 a (7:17) Substituting P0 .22 to represent the y measured values. yields following expression for the new u matrix.23.

the RLS solution is more robust than the standard least-squares solution (Equation 7. since such models can represent the data more accurately.22.2. Hence generating multiple piecewise linear (or curvilinear due to quadratic. A simple way of doing this is by comparing the Euclidean distance of the sample output color (assuming . which is identiﬁed by its centroid. compression. Convergence to the desired u is obtained when the error function is driven to zero or to a level below the desired threshold. Linear models are generated for each region.. The number of regions.13].24 into Equation 7. and signal partitioning applications [12. number of training colors in each region. It has been found suitable for classifying the multidimensional color space into smaller regions or clusters. L*a*b* values or CMYK values) to obtain partitioned colors and their centroids. When the adaptation is required. cubic terms) models at various input points and combining them via interpolation can lead to much more efﬁcient models than global linear models.25 and 7. where a large set of input color nodes are divided into groups having approximately the same number of points closest to them. A suitable cluster assignment process is required to assign each color sample to the right cluster. The piecewise linear approach to modeling nonlinearity of many engineering systems is a well-known concept. 7.g. Hence.1. the following recursive estimation formula emerges À Á ukþ1 ¼ uk þ Pkþ1 akþ1 ykþ1 À aT uk kþ1 where Pk þ 1 is Pkþ1 ¼ Pk À Pk akþ1 aT Pk kþ1 1 þ aT Pk akþ1 kþ1 (7:26) (7:25) Equations 7.26 achieve the adaptive estimation by using the previous estimate and the error between the outputs to converge to a new estimate. including printers. this recursive estimation can be initiated by simply setting P0 to a diagonal matrix. (2) cluster the input or output colors (e.6). and by letting u0 be the best ﬁrst guess.4. iteratively execute the RLS algorithm using Equation 7. only enters into the recursive equation in the ﬁrst guess for u0 (Equation 7. In practice.5) based on vector quantization—a classical quantization technique— widely used in the literature for a variety of classiﬁcation. A0. and the methodology required for partitioning the multidimensional color space depend on the device under study. It is related to self-organizing maps. The matrix.Three-Dimensional Control of Color Management Systems 313 Substituting Equation 7.25 for each of the clusters separately.3 Piecewise Linear Models In many practical imaging devices. and (3) obtain parameters of the linear model for each cluster using Equation 7. Key steps involved in constructing the initial multiple piecewise linear models based on clusters are (1) obtain training samples via experiments or historical database.4. To fully exploit the power of piecewise linear models.3. the linear model with the RLS algorithm cannot accurately ﬁt one uniform function to the entire input–output data set since the functional relationship between input–output data may be locally nonlinear or may be only partially linear. the color space is partitioned into a number of regions or clusters. K-means algorithm (discussed in Section 2.6).

described below.000 for 104 characterization samples. . . The number of elements could be higher when a reﬂectance spectral database is used for clustering (not considered in this example). 3. and [y11. Example 7. Hence. . . we used the notation. . The process for generating and adapting piecewise linear models over time is illustrated in Examples 7. N will become larger. . . .6 shows a ﬂowchart illustrating the method of determining the centroids. Vector Ck is the centroid of the L*a*b* values in a kth cluster. as the gamut becomes larger and more nonlinear. a mapping of the CMYK values to L*a*b* be denoted as a database U ¼ ½u1 u2 . uN are vectors containing input CMYK values for each color containing four values denoted by n y1. uNk k 2 RnÂNk À k Â Ã lÂNk ¼ Ck y11 y22 . . vectors u1. . each having three elements. K as follows: Â Ã !Y Uk ¼ u11 u22 . yNk k 2 R (7:28) where ½u1 u2 . uN k are the vector elements containing the Nk input CMYK values for the kth cell. which is a predetermined number based on the number of training samples that need to be classiﬁed. of Equation 7. depending on the training set. . y2. or 164 and so on. SOLUTION The method of classifying the color input–output data is described ﬁrst. The relationship between K and N is as follows: N¼ K X k¼1 Nk (7:29) Figure 7.1 Use K-means (Linde–Buzo–Gray [LBG]) algorithm [12] to divide the 104 input–output printer modeling data into 10 clusters.1 and 7. . Uk for k ¼ 1.28 because of the reordering of the training sample data that occur during the clustering process.314 Control of Color Imaging Systems: Analysis and Design output colors were used to create the clusters) with each centroid. u22. N could be 44. . These clusters will be generated by the K-means algorithm. 84. These centroids will become the centers of the respective clusters of . N ¼ 10. . run the RLS algorithm. .. . . of Equation 7. . Let the training vectors. . uN 2 RnÂN À Y ¼ ½y1 ! y2 . After making the right cluster assignment. yNk]k are the vector elements formed with L*a*b* values. . the training database is partitioned into K clusters. u2.2. Describe the critical steps in the algorithm. u2. u11. . . and then choosing the cluster whose centroid has the shortest distance to the sample output color. . y22. After assigning the colors to the clusters. Generally. . yN 2 RlÂN (7:27) where u1.. .. . . u2. .. uNk to describe new input vectors. . 2. . Using the K-means algorithm. yN are vectors containing output L*a*b* values for corresponding colors containing three values denoted by l In this example. . .27 do not necessarily correspond to the u1. N will be smaller for a calibrated printer.

ε = distortion threshold. then use training database vectors u1. yN are used for centroid and cluster assignments in L*a*b* space. . …. . Accumulate yi into AJ 3. K } Also calculate the centroids for CMYK space using the color indices for each cluster and store them Yes Is Dm–1 – Dm Dm < ε? No FIGURE 7. Compute average distortion Dm = Store new centroids {Ck. k =1. . . . 2. y2. . A2 .6 Flowchart of the centroid and cluster assignment in L*a*b* space. . …. Set m = 0 and D 0 > 1 (a large positive number = > 1000). K } = initial cluster centroids.Three-Dimensional Control of Color Management Systems 315 Initialize K = cluster size. If the same is required in CMYK space. y2. k = 1. . …. yN ] Update K partition sets A1. Compute J = arg min D = arg min d (yi. AK 1. 2. C 0 = {C k . Note: The training database vectors y1. A1 = A2 = …= AK = empty set i = 15 E =0 Enter N = number of training samples from the database [u1. Compute distortion E =E + Dmin where Dmin = minimum of d obtained in part (1) i=i+1 Is i = N ? No Yes 1. uN] [ y1. . …. uN. u2. Lk = number of vectors in Ak E N m = m +1 E =0 2. Ck ) k k where d is the Euclidian distance 2. Find the centroid of each cluster Σ A (i) k i=1 Lk Ck = Lk . …. u2.

and repeats the previous steps as shown in the ﬂowchart. . AK. and from the Euclidean distances d. including e. . i. Next. Thus. e is a distortion threshold. Ck Þ k k (7:30) Then.316 Control of Color Imaging Systems: Analysis and Design the partitioned database. These are the clusters of the database. expressed as a color vector. K is the number of clusters into which the database is to be partitioned. i. which indicates the maximum allowable distortion. The average distortion Dm is obtained by Dm ¼ E N (7:33) Then. . The centroids Ck may be arbitrary. For each training sample yi. that is. . K. m. m and i are simply iteration counters. . an updated cluster centroid Ck is determined for each cluster A1. yi is accumulated into AJ. A2. K. and may be initially set at 0 and 1. . the Euclidean distance d to each cluster centroid Ck is determined. When i ¼ N. Empty sets A1. The larger the parameter K is. Accumulation of the other training sample yi into the appropriate cluster AJ continues until all training samples have been collected into the appropriate clusters.9–7. D0 is an initial distortion setting. when all training samples have been accumulated. which will be ﬁlled with initial values and then updated until certain algorithm criteria are met. 2. This determination may be made by determining whether the following relation is satisﬁed: Dm-1 À Dm Dm e (7:34) If Equation 7. the better the results obtained from the model (see Figures 7. AK are established a priori. where k ¼ 1. which is initially set at 0. D0. . as described below. .11). but it will increase the processing time. and is set at an arbitrary large positive number. E is a distortion value. and E. beginning this time with the updated . This determination may be performed according to the following equation: Ck ¼ PLk i¼1 Ak (i) Lk (7:32) where Lk is the number of vectors in cluster Ak. . . Various initialization values are entered. Initial cluster centroids C0 are set equal to Ck. A2. it is determined whether distortion is within the distortion threshold e. the process sets E ¼ 0 and m ¼ m þ 1. D0. m. respectively. as deﬁned by the criterion associated with desired system performance. The cluster AJ having the minimum Euclidean distance is identiﬁed as follows: J ¼ arg min d ¼ arg min d ðyi . one centroid Ck is assigned to each empty set. the distortion E is determined by E ¼ E þ dmin (7:31) where dmin is the minimum distortion value d obtained in earlier steps.34 is not satisﬁed. Speciﬁcally. or may be set using a ‘‘best guess’’ based on previous experience. and E are values used in the algorithm. . such as 1000. .

. the process stores the new centroids C1. . y2. . . Ck) k k 2. that is. 2. as shown in Figure 7. .7.CK. AK 1. However.6 in CMYK space using training database vectors u1. These centroids are the ‘‘ﬁnal’’ centroids that will be used in the clusters for the adaptation process.7. .6. A1. a ﬁnal step of accumulating the training samples into the appropriate clusters will be performed. …. a critical color in the yellow region of the gamut (colors used for controlling the printing device) or colors from a document obtained by processing the pixels in the document where an accurate model is desired.Three-Dimensional Control of Color Management Systems 317 cluster centroids Ck. . Compute J = arg min D = arg min d (yi. …. k =1. Using N number of training samples. uN. …. But. when using the models. A2. . Repeat this process until all training samples are accumulated and stored into Initialize K = cluster size Ck . These centroids—in L*a*b* space—will be used during the adaptation process. cluster centroids obtained from some other approach can also be used here. A2. u2. . This can be computed using Equation 7. …. After the centroids are computed. yi is accumulated into the appropriate cluster. . the corresponding cluster centroids in CMYK space is required. the centroids may be ‘‘critical colors. the cluster having the centroid with the shortest Euclidean distance from the training sample. Accumulate yi into AJ Is i =N ? No i=i+1 Yes Stop FIGURE 7. AK and set i initially at 1.34 is satisﬁed. C2.6 are used in the initialization step of Figure 7.32 in CMYK space using color indices from the ﬁnal clusters or running the ﬂowchart of Figure 7.7 Flowchart of accumulating training samples into clusters in L*a*b* space.’’ for example. yN] Update K partition sets A1. . AK = empty sets i=1 Enter N = number of training samples and training samples from the database [u1. K = cluster centroids A1. A2. For example. . begin. create an empty set. The cluster centroids Ck from Figure 7. Again. When Equation 7. u2. uN] [y1. similar to Figure 7. . ….

318 Control of Color Imaging Systems: Analysis and Design FIGURE 7. Describe the key steps in assigning clusters and the parameter adaptation process for the piecewise linear model.2) is another factor that requires optimization.8 Ten cluster L*a*b* gamut (clustering was done in L*a*b* space).e.11.1. . In general. Clusters obtained in CMYK space is not shown. . Figure 7. As might be expected. clusters A1. cluster #6 had a minimum of 550 colors and cluster #8 had a maximum of 2577 colors. . Example 7. A2..9 through 7.8 illustrates pictorially the clustered color gamuts in L*a*b* space. since a higher number of parameters may model the noise by overﬁtting the data.2 Using RLS algorithm and training samples from Example 7. the parameter estimation becomes less accurate. The number of modeling parameters (i.9 show that the accuracy error (both mean and 95%) almost decreases monotonically with an increase in the number of clusters. SOLUTION The accuracy error with respect to the number of clusters of the piecewise linear afﬁne model with 13 parameters in each cluster is shown graphically in Figures 7. number of elements in the u matrix. (a) obtain a piecewise linear afﬁne model and (b) piecewise linear quadratic and cubic model for different clusters. the curves in Figure 7. if there are more training . When the simulation reached 10 clusters. These data were purposely produced for an uncalibrated printer. Show the model accuracy as a function of the number of clusters. Thus the trend is clearly in the right direction. AK. As the number of training colors assigned to a particular cluster reduces. Equation 7. .

5 4 3. 4.5 5 4.5 4 3.10 Error between experimental data and a piecewise linear quadratic and cubic model as a function of the number of clusters (computed for up to 10 clusters) for test colors for printer A with 104 characterization data.5 2 0 2 4 Mean 95% 6 8 10 12 14 16 18 20 Number of clusters 8.Three-Dimensional Control of Color Management Systems 18 16 14 12 10 8 6 4 1 2 3 4 5 6 7 Number of clusters 8 9 10 ΔE CIELab ΔE 2000 35 30 25 20 15 10 5 1 2 3 4 5 6 7 Number of clusters 8 9 319 Mean 95% Mean 95% 10 FIGURE 7.5 6 5.5 3 2. .11 Error between experimental data and a piecewise linear quadratic and cubic model as a function of the number of clusters (computed for up to 20 clusters) for test colors for printer B with 164 characterization data.5 5 Mean 95% 11 10 9 8 7 6 5 4 3 1 2 3 4 5 6 7 8 Number of clusters Mean 95% 9 10 FIGURE 7. 5.5 4 3.9 Error between experimental data and a piecewise linear afﬁne model as a function of the number of clusters for test colors.5 7 6.5 3 2.5 8 7.5 6 5.5 Mean 95% ΔE CIELab ΔE 2000 0 2 4 6 8 10 12 14 16 18 20 Number of clusters FIGURE 7.5 2 1 2 3 4 5 6 7 8 Number of clusters 9 10 ΔE CIELab ΔE 2000 7 6.5 5 4.

.12. The above approach was used in our simulations in Figures 7. we need to run following steps at a minimum: 1.25 and 7. Although increasing number of clusters improves the accuracy. to update the u matrix corresponding to the cluster of choice based on the cluster having the centroid with the shortest Euclidean distance. Determine which of the Euclidean distances in step #2 is the shortest. This is done in the cluster assignment block.26. which in turn can provide improvements to the model accuracy. The same block diagram can be used for predicting the L*a*b* values of test colors when their CMYK values are known. 4. Calculate the Euclidean distances from the current color sample to each cluster centroid in L*a*b* space. Now. it also increases the number of models and hence the total number of parameters required for modeling the entire color space. This is done in the cluster assignment block of Figure 7. Receive the L*a*b* values from the sensor (from the customer images or obtained by scheduling patches). Cluster centroid and parameter matrix u for each of the clusters are used during this process. The prediction process is described by the following steps: 1. Note: we do not use the CMYK space as in the prediction process. 2. Figure 7. samples available then it will be easy to increase the number of clusters. during the adaptation process. 4. 2. Receive the CMYK values for the prediction set. and the number of clusters is necessary. Calculate the Euclidean distance from the current color sample to each of the cluster centroids in CMYK space.10 for a group of test colors that is separate from the training set. Run the RLS algorithm.9 and 7. 3. Therefore. This becomes the cluster of choice for the color sample.320 Control of Color Imaging Systems: Analysis and Design CMYK values for target color Printing device Color sensing device + – Cluster assignment Parametric model for assigned cluster Error FIGURE 7. 3.12 shows a block diagram of the adaptation process. a trade-off between the number of model parameters. which then becomes the cluster of choice for the color whose CMYK values are to be estimated using the model. Equations 7. Calculate the L*a*b* values using the updated parameters of the model corresponding to the cluster in step #3.12 System block diagram showing the adaptation process with cluster assignment used to tune the parameters of the piecewise linear model. Determine which of the Euclidean distances in step #2 is the shortest.

A more advanced twodimensional (2-D) PCA method can be found in Ref. Section 3.1 PCA-Based Model in Spectral Space In this section. although the representation can be made with any multivariate input–output data samples [21]. for speciﬁc input CMYK nodes. Hotelling transform. which outputs 36 spectral reﬂectance values. it gives less redundancy and as good a representation of the system as possible with a few principal components. singular value decomposition.4. 7.Three-Dimensional Control of Color Management Systems 321 Thus.10). Many researchers [14–18] have attempted to decompose the spectra in terms of principal basis vectors and use the basis vectors to model the imaging device or use the PCA to predict the reﬂectance measurements on multiple substrates (media) while characterizing PCAs on a reference substrate [17]. as well as signiﬁcant complexity associated with the computation.4. In PCA-based modeling. A much better way to model the input–output data empirically is by using principal component analysis (PCA) (which is also known by many other names. whose measurements cannot be obtained easily because the printer may be scheduled for production jobs. etc. PCA does not help. This diagonal matrix is the matrix of singular values.. with different vector lengths) are not considered for PCA. Gretag spectrophotometer. the basis vectors become known as the most signiﬁcant or principal components. and an in-line spectrophotometer [7].2 PRINCIPAL COMPONENT ANALYSIS-BASED MODEL Though the piecewise linear models based on clustering can produce better models. empirical orthogonal function decomposition. a XRite spectrophotometer [10]. while at other times models can be used to estimate the L*a*b* values using the recently updated parameters of the updated proﬁle LUTs. evenly spaced at 10 nm over the visible spectrum over the spectrum 380–730 nm. This method is widely used in other ﬁelds of engineering [19. A standard spectral sensor is used for measurement. the implementation may suffer from a large body of parameters to track.2. When sorted from highest singular value to the lowest. Karhunen–Loeve decomposition.. In PCA. such as. It decomposes data into linear combinations of mutually orthogonal and basis vectors of unit length (extracted from a training data set) in such a way that in that basis. The PCA approach extracts key information from the data for modeling and throws away the rest. we ﬁrst show a one-dimensional (1-D) spectral PCA-based model for a static printing system. As a result of this operation on the input–output data. output data with a mixture of different wavelengths (i. it is essential that the spectral training data is mutually correlated.e.20] and may receive more use in color systems as the demand for high performance systems increase. A simple online estimation method is introduced later for a 1-D PCA to update the parameters during the adaptation process. Also. 7. we show the approach when the sensed output is the reﬂectance spectra. [22] which uses a signiﬁcantly . the second order statistics (covariance matrix) of the data is diagonalized. all of them nonnegative by deﬁnition. If they are independent. parameter updates are done adaptively based on sensor measurements to capture the current state of the device. which has 31 outputs evenly spaced at 10 nm over the wavelength of 400–700 nm. both time and computing resources.

. If R is the full reﬂectance spectra of the color. R0(l) is the sample mean. given a stream of output spectral training data. R1. W ¼ [W1 W2 Á Á Á WK]T. The vector. We show below how to model the printer using a 1-D PCA approach. which is computed from the output spectral data. n ¼ 31 when there are 31 reﬂectance values evenly spaced over the spectrum of 400–700 nm). 7 4 . The basis vectors are obtained as follows. 5 . . and Wj is the jth weight parameter. in matrix form. First compute the covariance matrix formed by the output spectral data X ¼ N 1 X ½Ri À R0 ½Ri À R0 T N i¼1 (7:39) . the terms inside the summation of Equation 7.322 Control of Color Imaging Systems: Analysis and Design smaller number of basis vectors. . is the parameter vector of size K Â 1 containing the scalar weights for the color. RN.. R0 (l) ¼ N 1 X Ri (l) N i¼1 (7:36) We can represent.35 as 3 W1 6 W2 7 6 7 r ¼ ½c1 c2 Á Á Á cK 6 . The matrix. then each of these spectra can be modeled as a random sample from a mixture model in terms of K basis functions where K < n as R(l) ﬃ R0 (l) þ K X j¼1 Wj cj (l) (7:35) where cj(l) is the jth basis function derived from PCA analysis.37 can be written in matrix form for each color as r ¼ BW (7:38) 2 (7:37) where r is the zero-mean reﬂectance spectra of size n Â 1 of the color. reduced computation and measurements for continuous adaptation. R2. B ¼ [c1 c2 Á Á Á cK]. where each Ri is a vector of length n (e. whose columns contain the basis vectors with elements at the wavelength intervals used in the output spectral data. then the approximated spectral data is given by R ﬃ R0 þ r which represents the full reﬂectivity vector for that color at speciﬁed wavelengths. is the mixture matrix of size n Â K.g. For N number of input colors. WK Equation 7. .

In Figure 7. if the color gamut of the device is large. while the rest are ignored as noise. In some situations.35) is shown when up to nine basis vectors are used. A linear weighted combination of all these basis functions represents the complete spectra. are obtained in various ways. SVD X ¼ X ¼ CP2 CT ¼ n X i¼1 P 2 c i cT i i (7:40) The vectors. Also. correspond to the ith basis function. In Figure 7. . jP1j > jP2j > Á Á Á > jPnj are rank ordered singular values contained in the matrix. as in a 6 or 14 color press.40.Three-Dimensional Control of Color Management Systems 323 Perform the singular value decomposition (Section 3. The PCA method assures that the principal directions determined by the spectral PCA stage are orthogonal to each other. dominant eigenvalues. P P1 6 0 6 P¼6 0 6 4 : 0 2 0 P2 0 : 0 0 0 P3 : 0 : : : : : 3 0 0 7 7 0 7 7 : 5 Pn (7:42) The parameters. .40 [23].13. . weights are computed from the following dot product. 2. K. The actual spectra from the training set along with the DE 2000 and DE CIE Lab numbers are shown in the same ﬁgure to compare with the approximated spectra. in Equation 7. That is R(l) ¼ R0 (l) þ n X j¼1 Wj cj (l) (7:41) Note that CT C ¼ I ¼ identity matrix.15. One straightforward way is by using the orthogonality property of the basis vectors emanating from Equation 7. ci. . then only few. whose gamuts are typically larger than four-color press. They are pairwise orthogonal and orthonormal just as the sine and cosine functions in a Fourier decomposition of the composite function. n. it is difﬁcult to know precisely how many dominant eigenvalues (hence eigen or basis vectors) are required for a good approximation. In general.14. Wj with j ¼ 1. the ﬁrst four basis vectors are plotted as a function of wavelength. In Figure 7. K < n.10) on the covariance matrix to get n number of basis functions. then more basis vectors are required for modeling the device.35. Wj ¼ r T cj (7:43) If there are only few. . Since the basis vectors are also orthonormal. basis vectors are needed to estimate the spectra as in Equation 7. an example of reconstructed spectra (Equation 7.

Once the basis vectors are selected. required for modeling the parameters. Y. . M. K) is modeled as fj (C. An adaptation estimation algorithm is sometimes more valuable for modeling the parameters. over 99% of the spectral energy is captured in the ﬁrst six basis vectors.1 –0. . K) ¼ bj þ a1j C þ a2j M þ a3j Y þ a4j K þ a5j CM þ a6j CY þ a7j CK þ a8j MY þ a9j MK þ a10j YK þ a11j C 2 þ a12j M 2 þ a13j Y 2 þ a14j K 2 þ Á Á Á (7:45) (7:44) For simplicity.4 0. .3 –0. As shown in Table 7. 2. K) The function fj (C. Y. M. the weights Wj are modeled in terms of input variables. M. CMYK.3 0. . For convenience. preferably by spanning a large portion of the operating space of the device. and K number of such set .2 0.1. using Equation 7. Equation 7. K.4 400 1st 2nd 3rd 4th Control of Color Imaging Systems: Analysis and Design 450 500 550 600 Wavelength (nm) 650 700 FIGURE 7. Wj ¼ fj (C.13 First four basis vectors as a function of wavelength.324 0. the DE numbers are plotted as a function of the number of basis vectors to decide how many basis vectors. 2. Y. Wj.45 is written in matrix form for a simple linear afﬁne model containing ﬁve parameters for each weight. K basis vectors and for i ¼ 1. . .43 for j ¼ 1. .1 0 –0. for this printer.2 –0. N colors. using the training samples. . the color subscript i is not shown in zero-mean reﬂectance r and Wj in the above equations.

4 0.15 500 550 600 650 700 Wavelength (nm) Wavelength (nm) Wavelength (nm) FIGURE 7.45 0.4 0.35 0.25 0.05 0 450 500 550 600 650 700 –0. (From left to right: top row: 1.2 0.2 Reflectance values Reflectance values Three-Dimensional Control of Color Management Systems 0.3 0.25 0.45 0.14 Reconstruction of the spectra using a sequentially larger number of basis vectors for a sample blue color. and 3 basis vectors.25 0.35 0.2 Reflectance values Reflectance values 0.05 700 0 400 Wavelength (nm) Reflectance values 0.45 0.2 0.4 0.1 0. 2.3 0.05 400 450 500 550 600 650 700 0.05 450 500 550 600 650 700 0 400 450 0.05 0 400 450 500 550 600 650 Wavelength (nm) 0.4 0.15 Wavelength (nm) 0.2 0.15 0.1 0.35 0.05 700 0 400 0.35 0.1 0. solid curve represents actual spectra and dashed curve represents reconstructed spectra).3 0.45 0.25 0.3 0.4 0. and 6 basis vectors.1 0.15 0. 5.0. bottom row: 4.25 0.2 0.35 0.25 0.05 0 400 450 500 550 600 650 Reflectance values 0.15 0.45 0.3 0.45 0. 325 .1 0.3 0.1 0.4 0.15 0.35 0.

35 0. * DE2000 and DEab as function of the number of basis vectors for training TABLE 7.28 8. of basis vectors No. etc. WT will be a matrix of size N Â K.41 2. the least-square solution to the parameter matrix is given by À ÁÀ1 u0 ¼ AT A0 AT W T 0 0 (7:47) T . of basis vectors FIGURE 7.14 2 ½W1 : : WK ¼ [1 W2 C M Y 6 6 a11 6 K]6 a21 6 6 4 a31 a41 b1 b2 a12 a22 a32 a43 : : : : : : : : : : 7 a1K 7 7 a2K 7 7 7 a3K 5 a4K bK 3 (7:46) W ¼ A0 u0 Where WT ¼ [W1 W2 .92 62.05 * DEab 60. and CMYK values with a suitable structure of the model (linear afﬁne. WK] is a row vector containing K number of weights for the ith color. Hence.95 21.15 colors. Equation 7.02 3.326 45 40 35 30 Control of Color Imaging Systems: Analysis and Design 120 100 80 95% Max Mean 95% Max Mean ΔE2000 * ΔEab 1 2 3 4 5 6 7 8 9 10 25 20 15 10 5 0 60 40 20 0 1 2 3 4 5 6 7 8 9 10 No.09 0.69 0. When N number of colors are used in the training set. and u0 contains the new parameter matrix with K number of columns. quadratic.46 is very similar to Equation 7.). cubic. A0 contains N number of rows.82 5.3.91 1.1 DE Numbers Are Shown to Indicate the Approximation Accuracy Number of Basis Vectors 1 2 3 4 5 6 DE2000 23. .

Use the zero-mean spectral r from step 4. . of the N training set. . at minimum. . .26. .2 PCA-Based Modeling for Adaptive Estimation For performing continuous adaptation as fresh samples become available. using Equation 7. .41 using weight vector from prediction process step 3 and basis vectors from the training process step 3. . we have the complete model of the printer. K. u0. R0. . a random subset (1=2 in this case) of the data should be used . an RLS algorithm (Equation 7. . . bootstrapping is often needed. . 2. Determine matrix WT ¼ [W1 W2 . A general warning is required while using PCA method for modeling color systems. Solve the least-square solution for the parameter matrix. i ¼ 1. WK] for all prediction colors (or Wj for j ¼ 1. RN. . R1. 4. by grouping CMYK values corresponding to each training sample in A0 for the appropriate model structure and forming the weight matrix. the subscript i is not shown in r.47). .43 for j ¼ 1. . When it is used in combination with the 1-D spectral PCA vectors.46 since the parameter matrix u0 is known from the training process of step 6 above. 6. Determine the sample mean. . . 2. In bootstrapping. . 3. 3. Form covariance matrix (Equation 7. To use the parameter vector for predicting the spectra for any new CMYK values. Arrange input color CMYK values in the matrix A0 for the structure chosen during training step 6. or to improve the accuracy of parameters instead of the least squares. For convenience. .47. WT from step 5.25) can be used with slight modiﬁcations as in Equation 7. N. 2. u0.40). ri ¼ Ri À R0. is a function of CMYK. Obtain predicted reﬂectance spectra from Equation 7. N colors. . 4. Receive spectral vectors. . . run following key process steps: 1. . Determine zero-mean reﬂectance spectra for each training spectra. and determine K number of basis vectors cj with j ¼ 1.39). using Equation 7. 5. . 2. 2.48. À T Á ukþ1 ¼ uk þ Pkþ1 akþ1 Wkþ1 À aT uk (7:48) kþ1 where the equation for Pk þ 1 is shown in Equation 7. In order to gain conﬁdence in the use of experimental data.36. .4. 7.2. K for each color) from Equation 7. Determine weights Wj using Equation 7. Receive the CMYK values at which spectral reﬂectance are to be calculated using the model. The training process is summarized in following key steps: 1. Color numbers are omitted in the matrix A0 for simplicity. 2. K basis vectors and for i ¼ 1.Three-Dimensional Control of Color Management Systems 327 The parameter matrix. . R2. perform singular value decomposition (Equation 7. Adaptive estimation is initiated by setting P0 to identity and initial u0 to the output of the least-squares equation (Equation 7. . 2.

97 41.63 25.04 77.2 for different number of basis vectors.37 25.31 41.79 41.68 7. 2000 prediction colors is shown in Table 7.78 Max 87.89 4.50 32.46 13. Example 7.65 17.81 87. Improvement to the prediction accuracy saturates when the number of basis vectors becomes equal to ﬁve for both algorithms.45 55.28 19. input CMYK values are sampled uniformly between 0 and 255 to create 104 input–output characterization data set. of Basis Vectors 1 2 3 4 5 10 20 1 2 3 4 5 10 20 DE2000 Mean 20.80 10.27 10.30 60.61 Mean 31. The u matrix is of size 23 Â 10 (when 10 basis vectors were used) with mean propagation.72 57.65 7.75 7.66 57.25 29.80 20.60 25.74 7.52 27.5 illustrates the bootstrapping method for an example dataset.20 10. Problem 7.97 42.52 RLS .99 45. TABLE 7.25 19.02 56.20 Max 51.52 16.81 25.71 * DEab 95% 67.65 7.59 4.36 45.44 57.74 7.79 19. Accuracy for approx.79 19.40 84.02 41. Plot the model accuracy for test colors as a function of the number of basis vectors.24 19.81 27.53 13.23 46. This process should be repeated multiple times. RLS algorithm is found to be more accurate than a simple least-squares regression.52 13.46 7.328 Control of Color Imaging Systems: Analysis and Design for the PCA to determine PCA vectors.25 66.19 4.25 and 7. The results of the PCA vectors are then applied to the rest of the data set to determine how good a model with PCA can be for the whole set.60 24. The spectral data for each patch is measured at 31 wavelengths from 400 to 700 nm at 10 nm intervals.79 47.58 95% 35.54 10.26. If all of the data come from one distribution.41 10. the expected amount of residual energy (actual variation–modeled variation) is insensitive to the subset chosen for analysis using PCA.50 25.46 13.83 19.66 7. How many basis vectors are needed to obtain a reasonably accurate model? SOLUTION We will set P0 ¼ I (Identity matrix) in the RLS Equations 7.90 13.85 19.89 56.22 57.43 29. Using PCA vectors ﬁnd parameters for a least squares and RLS model.34 60.16 29.21 42. Use spectral PCA mathematical techniques to analyze the characterization data and ﬁt a model.2 Model Accuracy Shown with Respect to Spectral PCA Basis Vectors Models Least squares No.58 4.28 34.94 17.27 19.71 7.16 17. All these calculations were done with absolute L*a*b* values. by choosing each time a different random subset for PCA analysis and the rest of the data to ﬁt with those PCA vectors.73 57.3 In a CMYK printing system.63 4.85 29.06 7.87 13.99 32.

except that the reﬂectance spectra are replaced by its natural log R(l) ! Àlog (R(l)) (7:49) A summary of the estimation algorithm for Log-PCA is shown in Figure 7.2. find weight vector using prediction process steps in Section 7.Three-Dimensional Control of Color Management Systems 329 7. .16a and b.1 Estimate-log of reflectance spectra X = –log(R(λ))= ΣWiψi (λ) i=1 8 Use steps described in Section 7. It requires a reasonably large Get training data (CMYK)1 R1 (λ) R2 (λ) RN (λ) .2. Example 7.3 in which the spectral data were not converted with log-PCA.3 shows the results for two different least-squares models. R(λ) = exp(–X) (a) (b) FIGURE 7.48 Obtain predicted reflectance spectra from equation. SOLUTION Table 7.4.1. since a few principal components can capture the essential characteristics of the printer. (b) Spectral prediction algorithm using model parameters and log-PCA.16 (a) Parameter training algorithm shown for log-PCA model.2. use Equation 7. clustered linearized PCA models can be generated since the clustering approach can model the nonlinearities well. (CMYK)2 (CMYK)N Form covariance matrix: Σ =[–ln(R)]T[–ln(R)] where R = [R1 R2 … RN]T Given CMYK. . For RLS algorithm.4. θ 0.3 Log-PCA Model (Log-PCA) The Log-PCA is sometimes used instead of the linear PCA described in Section 7. .4 Piecewise Linear PCA Model The main purpose of the PCA approach is to model the printer data set from a few principal components and a few weights (Equation 7. The approach is similar to linear PCA.4.4.3 input–output characterization data and apply log-PCA model to show the improvements to modeling accuracy.2. 7. To improve the accuracy of the printer model. Both algorithms give similar accuracy when compared to Example 7.2.43).4 Use Example 7.4.1 to estimate the parameter vector.

48 26. each having some speciﬁed number of elements (say 31 elements when the data is captured between 400 and 700 nm at 10 nm interval) and vector Ck is the centroid of the spectral values in a kth cluster.93 11.96 19.38 Max 90.68 61.75 7.43 23.330 Control of Color Imaging Systems: Analysis and Design TABLE 7.15 36.60 27.25 44.60 27.36 7.86 14.38 74.95 4.92 5.21 10.96 4.4.21 10.38 14.38 7. SOLUTION Results are summarized in Table 7. Example 7.90 14.82 26.48 26.5 Yule–Nielson Corrected PCA Model Yule–Nielson correction factor can sometimes be introduced to the measured spectral data from the printer before applying the PCA.94 36.38 7.92 5.00 10.38 14.68 10.3 Model Accuracy Shown with Respect to Log-PCA Basis Vectors Models Least squares No. conceptually the rest of the modeling procedure is very similar to the one described in the previous sections.36 * DEab 95% 74.82 26.43 5.68 10.5 A clustered PCA-based model is to be constructed for a 104 spectral characterization data.2. After clustering.36 7.21 14.02 14.90 40.96 95% 36.43 23.90 Mean 32.97 74. Cluster assignments are done in the output (i.00 10.23 7.25 44.33 4.e. Obtain spectral PCA vectors for each cluster.15 10.89 26.72 26.86 14.15 10. Example 7. Hence.97 74.79 14.89 90. This factor is somewhat related .90 14.68 61.66 7. If [y11 y22 Á Á Á yNk]k are the vector elements of the spectral sensor at speciﬁed wavelengths.82 14.96 4.4.06 14.82 14.4.75 7.66 7.49 14.21 14.49 14.94 36.15 Max 40. Clearly this model is more accurate when compared to Example 7.89 26.06 14.23 7. Model each cluster with RLS algorithm. Show the performance of your model for a test data as a function of the number of clusters.5 illustrates the performance improvement gained with this approach for a nonlinear printer.45 15.95 4.93 14.93 14.. L*a*b*) space. of Basis Vectors 1 2 3 4 5 10 20 1 2 3 4 5 10 20 DE2000 Mean 19. then K-means algorithm is one of the candidates available for us to cluster the data in spectral space. Use the K-means algorithm to cluster the color data in spectral space.43 5.33 4. we have not repeated the entire process again.89 RLS number of training samples to adequately capture PCA vectors within each cluster.02 14. 7.79 14.45 15.93 11.36 32. Use ﬁve or six PCA basis vectors.72 26.

3 NEUGEBAUER MODEL One well-known spectral modeling technique is the Neugebauer color mixing model [25–64] commonly used to model the output color in digital printing processes.49 Max 53.67 5.Three-Dimensional Control of Color Management Systems 331 TABLE 7.46 12.13 5.96 91.01 2.60 92. there are 8 (23 ¼ 8) Neugebauer primaries. black (CMYK) colorants. magenta.73 4.47 30.58 10.11 10.31 39.65 to the physical effects of light scattering and mechanical dot growth [24.68 6.98 7.89 19.95 22. The following sections illustrate various color models studied by many researchers.61 Mean 7. and more combinations of available colorants and a substrate on which the colorants are formed.56].23 * DEab 95% 23.92 2.76 3.23 32. models based on the fundamental physical process are much more preferred. if available.81 31. Often physics-based approaches would require fewer measurements. 7. except that the reﬂectance spectra are raised to the power of (1=m).14 3. successful printer characterization based purely on experimental data and its piecewise linear representation may need a large number of measurements.46 14. The Neugebauer model determines the spectral reﬂectance function in terms of the weighted sum of spectral reﬂectance functions obtained from one.07 3.23 4.0. Consequently.40 3.88 53.49 12. As clearly seen from previous discussions.32 7.47 6.30 20.96 57.94 9.04 127. having cyan.12 7. where m is determined by trial and error to achieve the best prediction. The resulting colorant combinations are referred to as the Neugebauer primaries. In an N color system.72 95% 10.64 12.4 * Accuracy of Clustered Spectral PCA Model (DE2000 and DEab Space) Models RLS No.29 6. In mathematical form.39 100. of Basis Vectors 1 2 3 4 5 6 7 8 9 10 DE2000 Mean 4.39 137.89 4.73 102. for example.00 41. the vector corrected . yellow. In a four-color system.62 49.02 4. having cyan. The approach is similar to normal PCA. there will be 16 (24 ¼ 16) Neugebauer primaries.74 2.79 Max 162.11 3.10 6.25 7. magenta. It can also exceed 2.98 4. R(l) ! R(l)1=m (7:50) Values of m are somewhat arbitrary and could be set anywhere between 1 (for a glossy substrate) and 2 (for a perfect diffuser).33 16. In a three-color system.23 11.51 7. there will be 2N known spectral reﬂectance functions. two.74 50.51 5. for example.57 20.4. yellow (CMY) colorants.52.82 7. since the nonlinearities of the physical process can be captured in the analytical functions.

YK. CK. can be tuned from spectral reﬂectance measurements using standard least square or RLS algorithms as described in earlier sections. and vector W ¼ [W1 W2 .g. . Wi. the Neugebauer equation (Equation 7. WK]T is the Demichel weight vector. r ¼ R1=m. Alternatively. ri ¼ Ri . and a free parameter m is the Yule–Nielson correction factor The penetration represented by the m factor was expressed as a function of wide band reﬂectance. MY. . K. . . CMY. let us denote the Yule–Nielson corrected ith reﬂectance values at a 1=m given wavelength by the vector.51. CM. Y. With the corresponding new notation for the estimated spectra. 7. CY. is the matrix of size n Â K (e. CMK. MYK. They are the measured reﬂectance spectra of the corresponding primary samples with 100% area coverage on paper white: Ri (l) ) Reflectance Spectra of {W. B ¼ [r1 r2 . Wi. . Equation 7. K.4. C. .3.332 Control of Color Imaging Systems: Analysis and Design Neugebauer model for a CMYK printer with corrections to the penetration and scattering of light onto paper [52–56] is given by the following equation " R(l) ¼ K X i¼1 #m Wi Ri (l) 1 m (7:51) where R(l) is the average output spectral reﬂectance (estimated) Wi is the Demichel weighting of the ith Neugebauer primary Ri(l) is the reﬂectance spectra of the ith Neugebauer primary which is the ith basis vector. and tuning of Neugebauer primaries with least squares or weighted least squares using training spectral reﬂectance measurements are other basic ideas being attempted for CMYK printers [30]. K is equal to the number of Neugebauer primaries. The weights can be further represented as individual fractional area coverages as in Demichel’s equation [56–58] and then these fractional area coverages can be tuned from spectral measurements. n ¼ 31) when there are 31 reﬂectance values. ﬁxing the weights. CMYK} Weights.48]. assumes that the reﬂectance of a spatial area is the additive combination of the reﬂectances of the primary colors and their overprints. . . 2. the m factor has been expressed in terms of narrow-band spectral curves and is allowed to vary over a wide range to obtain a best ﬁt with the measurement [47. More recently. MK..51) can be represented in vector form as r ¼ BW (7:52) The matrix. . where i ¼ 1. The Neugebauer model. rK]. which is 16 for a four-color CMYK printer. CYK. M.1 Parameterized Model for Neugebauer Weights For convenience.

The new parameter matrix u0 can be estimated using the least square Equation 7.47 or RLS Equation 7.Three-Dimensional Control of Color Management Systems 333 With a few matrix operations. develop a spectral Neugebauer model (Equation 7. 30 25 20 ΔE2000 15 10 5 0 1 1. Demichel weights are modeled with respect to input CMYK values using least-squares algorithm with linear afﬁne and quadratic models. A trial and error approach can be adopted to come up with a best value for m. and m ¼ 2 corresponds to Lambertian or perfectly diffused scattering in the paper [31].47.6 For a four-color CMYK printer. SOLUTION When Yule–Nielson factor m is equal to 1.51) with Demichel weights for 16 Neugebauer primaries. b. a. For an experimental printer considered in this simulation.18 shows the model accuracy as a function of the number of uniformly sampled data sets for best value of m from Figure 7.5 2 2.5 3 Yule–Nielson factor Mean 95% * ΔEab 3. 54.17 shows the model accuracy as a function of parameter m and Figure 7. there is no light scattering in the paper. 44. we used Bayer’s dithering technique to produce a halftone patch. Example 7.44 through 7. etc.48. As the halftone screen changes. Plot the accuracy of the model as a function of m.17. . we can express the weights as a function of the primaries and the r vector as follows: À ÁÀ1 W T ¼ r T B BT B (7:53) The weight vector W can now be modeled in terms of a new parameter matrix u0 (with linear afﬁne=quadratic=cubic model parameters) and input variables. CMYK. 64. 34. Estimate the weights using training samples with least-squares algorithms and Yule–Nielson factor (m) as a free parameter. the optimal m could be different. For the best m..5 4 FIGURE 7.g.5 2 2.5 3 Yule–Nielson factor 3. plot the model accuracy as a function of the number of uniformly sampled data sets (e.).17 The accuracy of the Neugebauer model for IT8 colors plotted as a function of the Yule–Nielson factor. Figure 7.5 4 Mean 95% 90 80 70 60 50 40 30 20 10 0 1 1. using the training samples as in Equations 7.

60 4.24 18.51 9.34 18.93 36.91 69.32 69.83 51.79 18.54 b* 28.10 72.98 7.03 18.36 b* 41.03 10.40 31. of samples No.45 9.10 16.39 17.10 16.87 6.99 16.15 15.30 2.60 4.36 3.32 68.22 Max 32.73 69.14 16. Since nonlinearities of the physical process are captured in the function.91 50.28 Max Color a* À5.5).76 9.334 11 10 9 Control of Color Imaging Systems: Analysis and Design 20 18 16 Mean 95% Mean 95% ΔE2000 7 6 5 4 101 102 103 * ΔEab 104 8 14 12 10 8 6 101 102 103 104 No.24 * DEab No.94 74.06 38.96 2. the model gives reasonably good prediction within a few hundred training samples (Table 7.33 18. the accuracy of Neugebauer models cannot be improved beyond a certain number based on increased training samples.35 9.86 0.76 72.51 17.98 5.11 7.31 66.69 67.07 18.22 69.72 15.33 33.14 L* 18.62 17.10 4.68 16.71 L* 65.95 68.5 * Accuracy of Neugebauer Models (DE2000 and DEab Space) DE2000 No. of Samples 81 256 625 1296 2401 4096 6561 10000 Mean 7.25 4.77 51.74 53. TABLE 7.15 32.61 9.24 4.30 17.07 7.86 17.17 95% 10. Evidently.27 9.18 The accuracy of Neugebauer model plotted as a function of the number of characterization samples.19 4.07 17.11 4.94 67.58 52.44 16. of samples FIGURE 7.98 7.82 . of Samples 81 256 625 1296 2401 4096 6561 10000 Mean 4.59 7.99 7.12 6.33 4.50 Max Color a* À3.21 68.82 2.30 36.66 4.23 1.92 50.19 Max 15.06 95% 18.15 51.14 2.19 68.29 7.96 6.03 16.02 33.19 9.

3. m. m. The mappings from the input digital counts C. and K. y. where the c. Equation 7. and k are the actual fractional areas covered by cyan. and ai (with i ¼ 1. Another commonly used halftone conﬁguration is the dot-on-dot screen. M. K ¼ 5). multiple colors are achieved by varying dot area coverages of the primary colors. These areas are functions of the input digital counts (control values) C.4. 4) represent the printer colorants of increasing dot area coverage. and a need for reduction in number of training samples due to cost reasons. These steps are not always necessary unless there is a requirement for further tuning of the Neugebauer weights to improve the prediction accuracy of the model. For this screen design. the colorants are drawn with a decreasing area coverage [29.51 (with Equation 7.43]. (a4 À a3 ). Rp4 (l). the Demichel weights corresponding to the primaries can be obtained by a probabilistic model as W1 ¼ (1 À c)(1 À m)(1 À y)(1 À k) W2 ¼ c(1 À m)(1 À y)(1 À k) W3 ¼ (1 À c)m(1 À y)(1 À k) W4 ¼ (1 À c)(1 À m)y(1 À k) W5 ¼ (1 À c)(1 À m)(1 À y)k W6 ¼ cm(1 À y)(1 À k) W7 ¼ c(1 À m)y(1 À k) W8 ¼ c(1 À m)(1 À y)k W9 ¼ (1 À c)my(1 À k) W10 ¼ (1 À c)m(1 À y)k W11 ¼ (1 À c)(1 À m)yk W12 ¼ cmy(1 À k) W13 ¼ cm(1 À y)k W14 ¼ c(1 À m)yk W15 ¼ (1 À c)myk W16 ¼ cmyk (7:54) where c. 2. (a3 À a2 ). Rp2 p3 p4 (l). magenta. 4) represent the corresponding dot area coverages. if pi (with i ¼ 1. The ﬁve primaries consist È the corresponding measured reﬂectance spectra and of É are denoted by Ri (l) 2 Rp1 p2 p3 p4 (l). 1 À a4 g: (7:55) . Rp3 p4 (l). m. and k are called dot growth functions or dot area functions (see Example 7. Y. and K to the fractional area coverage values c. M.2 Dot Area Coverages and Neugebauer Weights For binary printers. and k. Y. Thus. In this system.54. Y ! y. Demichel’s equations [56. and black colorant toner dots. The weights corresponding to the primaries are expressed as Wi 2 fa1 . since the separations are printed independently and the relative position of dots is random.3. and the dot growth functions C ! c.7).57] can be used to express the Demichel weights as the primary dot area coverages.. M ! m.19 shows a schematic input– output diagram of the printer model with the weighted Neugebauer Equation 7.Three-Dimensional Control of Color Management Systems 335 7.54) represents a fourth order polynomial in c. Rw (l) . y.51. for a ﬁxed l. For the case where the dot locations for the colorants are statistically independent.51 contains ﬁve primaries (i. m. y. 2. and K ! k. 3. For random mixing. then the Neugebauer Equation 7. the Demichel mixing Equation 7. Figure 7. respectively. (a2 À a1 ).e. yellow. which are integers between 0 and 255. y. and k dots are placed at the same screen angle and phase as illustrated in Figure 7.20 for an ideal dot pattern with no noise in a four-colorant system.

54 a is a weighting parameter between 0 and 1 that is used to relatively weight the two models The interested reader may refer to the original work on this subject in Ref. In order to use the Neugebauer model to model the printer accurately. The ideal dot-on-dot mixing model does not consider noise and misregistration effects.20 Dot-on-dot halftone screen example. a combination of dot-on-dot and random mixing model is proposed [31] to improve the model accuracy by giving relative weights to the reﬂectance predicted by the dot-on-dot model and the random mixing model.55 Rr(l) is the spectral reﬂectance predicted by the random mixing model using Equation 7. Therefore.51. [31]. K Y M C FIGURE 7.336 Control of Color Imaging Systems: Analysis and Design W1 W2 W3 W4 W5 W6 W7 W8 W9 W10 W11 W12 W13 W14 W15 W16 C Cyan dot growth function c M Magenta dot growth function Yellow dot growth function m Mixing equations y Neugebauer equation R(λ) Y K Black dot growth function k FIGURE 7. The predicted spectral reﬂectance in the combined model is given by R(l) ¼ (1 À a)Rd (l) þ aRr (l) (7:56) where Rd(l) is the spectral reﬂectance that is predicted by the dot-on-dot model using Equation 7.19 Block diagram of the Neugebauer model for a color printer. the weights Wi are modeled differently based on the way the dots are positioned. While the basic form of the mixing equations is similar to Equation 7. we need to estimate the .

Spectral weights can be selected as w(l) ¼ max ((l). Other weighting functions are explored in Ref. (l). this gives N o 1 1 2 @J @ Xn ¼ ½RCi ðlk Þm À½ci PC ðlk Þ þ ð1 À ci ÞPW ðlk Þm ¼ 0 @ci @ci k¼1 (7:59) . RLS. . i ¼ 1. cyan) and the same method is carried out for other colorants to ﬁnd their respective dot area coverages. (l) are tristimulus functions (i. is shown for one of the separations (e..).Three-Dimensional Control of Color Management Systems 337 dot area coverages from the digital counts CMYK as well as the Yule–Nielson correction factor m. reﬂectance spectra of the primaries.e. As an example. 7. where (l). total least square [43]. which can be obtained by solving an optimization problem in (a) spectral space or (b) L*a*b* space.4. the derivative of J with respect to ci is set equal to zero. .3. Use these values in Equation 7. For w(l) ¼ 1. 7. RCi(l) is the estimated reﬂectance obtained using the Neugebauer model and is given by the following equation ^ RCi (l) ¼ ci PC (l) þ ð1 À ci ÞPW (l) (7:58) where PC(l) and PW(l) are the reﬂectance spectra of cyan primary and paper white. we show how to model the dot area coverages for a random halftone screen with the least-squares algorithm and the RLS algorithm.51 along with the measured Neugebauer primaries (i. CMYK. etc. To improve the modeling accuracy. To estimate the dot area coverage for cyan in a least squared error sense in the spectral reﬂectance space.57.3 Estimation of Dot Area Coverages Using Least Squares An approach of relating the dot area coverage to the corresponding digital counts. The dot area coverages may be determined from spectral reﬂectance samples using a parameter ﬁtting or any of the estimation algorithms (least square. The relationship between the dot area coverage and digital counts is called dot growth function. To minimize the error metric in Equation 7. the following metric is minimized J¼ N Xn k¼1 Â Ã 1 o2 1 ^ w(l)½RCi ðlk Þm À RCi ðlk Þ m (7:57) ^ In Equation 7.3. 2. color matching x y z x y z functions: see Appendix A).. M ¼ Y ¼ K ¼ 0.3. (l)..4. neural networks.1 Optimization in Spectral Space Consider a set of cyan reﬂectance measurements RCi(l) corresponding to input digital values C ¼ i. . . [31]. Ri(l).57.e. a wavelength dependent weight w(l) is often included to emphasize the errors in the regions of the visible spectrum to which the human visual system is most sensitive.g. 16 for four-color CMYK printer) to predict the spectral reﬂectance for any CMYK input. respectively. genetic algorithm. robust estimation [44]. (l)).

1.6 at 10 nm interval between 400 and 700 nm. .3.7) and then plotting CMYK values against respective digital counts CMYK. a closed-form least square solution is intractable and thus a simple numerical iterative solution is preferred. we obtain on o PN n 1=m À½PW ðlk Þ1=m ½RCi ðlk Þ1=m À½PW ðlk Þ1=m k¼1 ½PC ðlk Þ ci ¼ o2 PN n 1=m À½PW ðlk Þ1=m k¼1 ½PC ðlk Þ i ¼ 0.59 for the unknown ci. magenta. The overall error metric. ac ¼ am ¼ ay ¼ ak ¼ 0.g. Unlike in the previous approach. m(i). and neutral patches are printed and their L*a*b* values are measured. since the transformation from spectral domain to color space L*a*b* is nonlinear. Let DEC. DE(i) ¼ ac DEc (i) þ am DEm (i) þ ay DEy (i) þ ak DEk (i) þ ag DEg (i) (7:61) where the weights ac.338 Control of Color Imaging Systems: Analysis and Design Solving Equation 7. The modiﬁed equation is given by PN n ci ¼ k¼1 on o ½wm ðlk ÞPC ðlk Þ1=m À½wm ðlk ÞPW ðlk Þ1=m ½wm ðlk ÞRCi ðlk Þ1=m À½wm ðlk ÞPW ðlk Þ1=m o2 PN n m 1=m À½wm ðlk ÞPW ðlk Þ1=m k¼1 ½w ðlk ÞPC ðlk Þ (7:60b) 7. A dot growth function (curve) can be easily obtained by using single separation step wedges at discrete digital counts (see Example 7. as in the previous section.4 for more accurate prediction along the neutral axis rather than along individual colorant axes). DEY. yellow. DEM. Thus. and neutral patches from the Neugebauer model with the appropriate mixing equation. . the error minimization is carried out in the L*a*b* space to assign visually meaningful weights. and ag are chosen based on region of the color space where more accurate reproduction is desirable (e. y. 2.4. For case w(l) 6¼ 1 Equation 7. is deﬁned as the weighted sum of the above CIELab errors. a set of N cyan. yellow. y(i).2 Optimization in L*a*b* Color Space In this approach. The relationship between . Assuming that the dot growth functions are separation independent.60a is modiﬁed by multiplying each reﬂectance spectra by wm (l). ak. 255 (7:60a) The above process is repeated for other three colorants m. DEk and DEg be the CIELAB error between the measured L*a*b* values and their corresponding predicted values for cyan. .3. am. and ag ¼ 0. ay. DE.7 A four-color CMYK printer has 16 usage basis colors for combinations of two states (0% and 100% area coverages) whose measured spectral curves are shown in Table 7.15.. Example 7. and k(i). magenta. black. and k. . The above metric is minimized with respect to the dot areas c(i).

60.7 using a. Then the spectra are calculated using basis vectors from Table 7. m. Weights for the cellular model are obtained using Demichel weights. y. and synthesized spectral curves from Table 7. Obtain the dot growth functions showing the relationship between the digital input C. k) is modeled by Demichel dot model. and Equation 7.51. k (0 c.Three-Dimensional Control of Color Management Systems 339 weights of the Neugebauer primaries to the dot area coverages (c. y. M.e. Intermediate area coverages. dot areas for intermediate digital counts are obtained through linear interpolation. A much ﬁner division of the cube can be obtained by using a larger number of grid points.. k 1) for measured spectral samples shown in Table 7. For these grid points. which presumably can lead to improved accuracy.3. The least square error minimization (Equation 7. CMYK values) corresponding to their digital counts (i. Equation 7. The weighted least square error minimization (Equation 7. To further improve the dot area coverage solution. To complete the function. Figure 7.22 shows the structure of the cube for a cellularized Neugebauer model with grid points also having 50% area coverage combinations. K and the corresponding dot area coverages c.57) led to a closed form optimal solution for determining the dot area coverages (i.60b) SOLUTION Figure 7. y. Equation 7.60a) b. such as 50%. Given the digital counts (CMYK). CMYK values). were not included.4 Cellular Neugebauer Model (Lab-NB) The basic Neugebauer model.4. used basis vectors that are combinations of the spectral curves for primary colors with 0% or 100% area coverages and their over prints. and the surface reﬂectance.19. the ink transmissions. the Neugebauer primaries will increase from 23 ¼ 8 to 33 ¼ 27. Y. this operation is done with one important assumption: each dot area coverage value for a given separation is independent of other three separations. m.6. we need to know their digital counts and follow the Neugebauer calculations shown in the block diagram of Figure 7. 7.21 shows the dot growth functions constructed for each separation using the least-squares error minimization formula. Other physics-based models considered in printing systems are Clapper–Yule [38. or Equation 7.e. these weights.. . To accommodate the use of mixed color samples.54 for random screens. All the dot area coverages are substituted in the weight Equation 7. These area coverages form the grid points of a cube when CMY separations are involved. The usual ideas of error minimization (Equation 7. Equation 7. which introduces correction for errors by modeling the internal scattering.65] model. However.7. m. we use the individual dot growth functions to ﬁnd their corresponding dot area coverages.55.51. To predict the spectral reﬂectance for any new color. the error minimization steps should be performed with a mixture model so that the factors involved with multiple separations are also taken into consideration.54. samples with mixed colors have to be used while calculating the dot growth functions.55 for dot-on-dot screen or using various least-squares regressions.

04 0.01 0.02 0.02 0.02 0.73 0.01 0.30 0.01 0.01 0.02 0.01 0.86 0.03 0.03 0.02 0.04 0.03 0.02 0.74 0.01 0.31 0.03 0.02 0.01 0.82 0.01 0.08 0.01 0.02 0.01 0.01 0.03 0.05 0.02 0.01 0.02 0.06 0.01 0.01 0.04 0.01 0.01 0.02 0.01 0.06 0.15 0.03 0.04 0.39 0.04 0.02 0.02 0.01 0.89 0.01 0.04 0.16 0.02 0.01 0.01 0.50 0.06 0.03 0.01 0.02 0.02 0.02 0.29 0.50 0.47 0.03 0.02 0.81 0.01 0.85 0.02 0.06 0.01 0.02 0.32 0.06 0.01 0.05 0.01 0.84 0.23 0.06 0.90 0.01 1.01 0.06 0.76 0.01 0.02 0.01 0.07 0.02 0.11 0.02 0.06 0.07 0.01 0.04 0.71 0.01 Control of Color Imaging Systems: Analysis and Design White C M Y K CM CY CK MY MK YK CMY CMK CYK MYK CMYK 0.02 0.86 0.03 0.73 0.02 0.63 0.85 0.06 0.05 0.02 0.08 0.01 0.67 0.90 0.91 0.01 0.01 0.91 0.05 0.03 0.74 0.01 0.90 0.89 0.04 0.01 0.01 0.02 0.01 0.38 0.11 0.62 0.01 0.02 0.01 0.86 0.01 0.06 0.01 0.90 0.44 0.89 0.06 0.10 0.01 0.02 0.02 0.84 0.85 0.03 0.43 0.05 0.01 0.01 0.02 0.02 0.06 0.56 0.35 0.02 0.02 0.91 0.01 0.05 0.01 0.83 0.02 0.08 0.01 0.03 0.01 0.23 0.03 0.07 0.01 0.01 0.02 0.06 0.04 0.02 0.01 1.02 0.14 0.02 0.04 0.32 0.39 0.03 0.03 0.01 0.03 0.07 0.01 0.04 0.02 0.90 0.82 0.01 0.08 0.02 0.03 0.02 0.02 0.01 0.03 0.03 0.01 0.02 0.02 0.06 0.14 0.05 0.97 0.07 0.05 0.02 0.11 0.02 0.09 0.01 0.04 0.36 0.02 0.03 0.01 0.03 0.03 0.03 0.05 0.22 0.04 0.01 0.04 0.06 0.01 0.01 0.84 0.64 0.05 0.02 0.02 0.07 0.03 0.01 0.03 0.02 0.02 0.01 0.84 0.01 0.02 0.01 0.10 0.01 0.32 0.01 0.25 0.81 0.02 0.01 0.89 0.02 0.07 0.03 0.07 0.03 0.49 0.02 0.01 0.01 0.02 0.79 0.02 0.89 0.01 0.49 0.08 0.02 0.10 0.93 0.08 0.21 0.61 0.01 0.02 0.08 0.01 0.02 0.11 0.02 0.03 0.47 0.04 0.07 0.05 0.06 0.02 0.02 0.07 0.01 0.02 0.05 0.01 0.11 0.08 0.03 0.01 1.02 0.26 0.91 0.02 0.06 0.18 0.86 0.02 0.09 0.01 0.37 0.87 0.93 0.86 0.01 0.01 0.11 0.06 0.02 0.42 0.01 0.01 0.85 0.83 0.40 0.03 0.09 0.01 0.07 0.02 0.13 0.07 0.02 0.86 0.82 0.04 0.02 0.85 0.05 0.01 0.01 0.01 0.63 0.02 0.37 0.04 0.27 0.02 0.02 0.07 0.06 0.02 0.06 0.02 0.83 0.02 0.86 0.05 0.05 0.02 0.01 0.02 0.07 0.05 0.85 0.81 0.01 0.01 0.01 0.80 0.06 0.01 0.01 0.05 0.04 0.02 0.02 0.84 0.05 0.01 0.03 0.07 0.02 0.01 0.84 0.01 0.92 0.02 0.02 0.01 0.95 0.03 0.02 0.04 0.89 0.01 0.02 0.02 0.03 0.6 Reﬂectance Spectra of Neugebauer Primaries Reﬂectance Spectra between 400 and 700 nm at 10 nm Interval Primaries 0.01 0.01 0.02 0.07 0.03 0.08 0.01 0.90 0.05 0.04 0.07 0.03 0.01 0.72 0.03 0.91 0.02 0.06 0.03 0.06 0.77 0.03 0.16 0.02 0.68 0.01 0.46 0.05 0.01 0.01 .02 0.01 0.04 0.340 TABLE 7.16 0.72 0.04 0.33 0.03 0.08 0.03 0.02 0.02 0.01 0.48 0.05 0.01 0.01 0.08 0.02 0.84 0.03 0.01 0.01 0.07 0.01 0.04 0.05 0.02 0.02 0.02 0.02 0.01 0.82 0.06 0.06 0.07 0.02 0.01 0.89 0.09 0.01 0.04 0.16 0.01 0.01 0.02 0.21 0.44 0.02 0.03 0.01 0.90 0.01 0.89 0.

28 0.64 0.74 0.43 0.86 0.80 0.83 0.45 0.76 0.89 0.89 0.99 0.89 0.64 0.65 0.06 0.05 0.05 0.89 0.43 0.28 0.84 0.88 0.75 0.87 0.87 Digital Count C 28 57 85 113 142 170 198 227 255 0.24 0.66 0.37 0.83 0.88 0.46 0.69 0.25 0.28 0.40 0.76 0.87 0.83 0.79 0.71 0.42 0.86 0.88 0.11 0.89 0.86 0.90 0.59 0.86 0.87 0.44 0.08 0.TABLE 7.26 0.53 0.73 0.72 0.72 0.89 0.86 0.93 0.88 0.60 0.88 0.69 0.74 0.87 0.06 0.71 0.96 0.74 0.82 0.87 0.74 0.55 0.48 0.50 0.77 0.54 0.44 0.69 0.85 0.83 0.70 0.90 0.87 0.68 0.91 0.84 0.29 0.37 0.76 0.59 0.49 0.42 0.88 0.83 0.47 0.83 0.88 0.32 0.80 0.67 0.84 0.88 0.88 0.48 0.61 0.79 0.29 0.61 0.83 0.88 0.45 0.79 0.84 0.84 0.40 0.24 0.82 0.89 0.74 0.28 0.91 0.15 0.88 0.38 0.47 0.87 0.29 0.76 0.24 0.13 0.70 0.94 0.96 0.77 0.56 0.33 0.69 0.44 0.23 0.63 0.22 0.63 0.64 0.73 0.07 0.89 0.80 0.72 0.87 0.92 0.12 0.89 0.89 0.87 0.92 0.97 0.87 0.85 0.87 0.90 0.29 0.23 0.44 0.86 0.45 0.55 0.84 0.89 0.61 0.74 0.26 0.71 0.16 0.07 0.79 0.42 0.88 0.43 0.26 0.47 0.59 0.84 0.12 0.88 0.86 0.90 0.68 0.36 0.73 0.50 0.87 0.35 0.54 0.97 0.49 0.73 0.90 0.23 0.30 0.74 0.59 0.05 0.28 0.50 0.59 0.86 0.85 0.97 0.66 0.71 0.48 0.90 0.05 0.86 0.49 0.60 0.86 0.28 0.27 0.85 0.77 0.60 0.63 0.59 0.79 0.33 0.81 0.89 0.32 0.11 0.69 0.79 0.74 0.77 0.89 0.27 0.45 0.36 0.63 0.64 0.84 0.88 0.24 0.85 0.73 0.86 0.87 0.16 0.48 0.87 0.85 0.84 0.83 0.18 0.33 0.61 0.89 0.59 0.89 0.74 0.76 0.78 0.54 0.54 0.74 0.77 0.87 0.50 0.54 0.79 0.89 0.78 0.77 0.74 0.83 0.41 0.77 0.46 0.82 0.33 0.26 0.74 0.59 0.68 0.88 0.53 0.83 0.86 0.83 0.74 0.60 0.33 0.83 0.71 0.45 0.27 0.79 0.55 0.89 0.53 0.95 0.59 0.46 0.13 0.86 0.46 0.42 0.60 0.36 0.52 0.82 0.69 0.86 0.95 0.61 1.77 0.55 0.19 0.78 0.33 0.27 0.92 0.95 0.24 0.67 0.89 0.28 0.94 0.82 0.69 0.23 0.42 0.85 0.77 0.27 0.01 0.83 0.27 0.14 0.44 0.55 0.61 0.44 0.83 0.85 0.14 0.60 0.28 0.84 0.41 0.77 0.06 0.77 0.67 0.68 0.89 0.87 0.86 0.89 0.79 0.36 0.88 0.24 0.81 0.52 0.70 0.89 0.89 0.85 0.80 0.41 0.89 0.53 0.89 0.81 0.21 0.61 0.61 0.88 0.81 0.83 0.75 0.64 0.86 0.86 0.77 0.38 Y 28 57 85 113 142 170 198 0.32 0.68 0.79 0.75 0.28 0.54 0.08 0.87 0.50 0.86 0.88 0.56 0.83 0.56 0.24 0.54 0.53 0.86 0.87 0.57 0.89 0.50 0.85 0.28 0.29 0.55 0.87 0.24 0.88 0.90 0.89 0.42 0.43 0.76 0.12 0.78 0.91 0.87 0.54 0.49 0.85 0.13 0.87 0.39 0.82 0.43 0.86 0.74 0.81 0.53 0.84 0.74 0.81 0.06 0.27 0.32 0.86 0.81 0.85 0.38 0.7 Reﬂectance Spectra for Obtaining Dot Growth Function Reﬂectance Spectra between 400 and 700 nm at 10 nm Interval 0.31 0.90 0.16 0.73 0.57 0.48 0.73 0.86 0.39 0.62 0.89 0.75 0.54 0.78 0.80 0.90 0.30 0.42 0.89 0.89 0.84 0.53 0.53 0.86 0.90 0.88 0.33 0.83 0.38 0.83 0.06 0.12 0.58 0.81 0.89 0.88 0.79 0.87 0.87 0.26 0.85 0.86 0.88 0.88 0.84 0.61 0.86 0.70 0.80 0.83 0.84 0.89 0.81 0.25 0.19 0.22 0.36 0.88 0.28 341 (continued ) .99 0.89 0.87 0.14 0.87 0.24 0.95 0.88 0.29 0.49 0.30 0.79 0.90 0.61 0.14 0.16 0.88 0.80 0.92 0.99 0.68 0.96 0.89 0.25 0.06 0.94 0.45 0.83 0.42 0.76 0.92 0.98 0.34 0.85 0.88 0.26 0.78 0.85 0.87 0.85 0.37 0.89 0.43 0.87 0.76 0.73 0.05 0.05 0.66 0.32 0.87 0.23 0.81 0.40 0.79 0.81 0.76 0.90 0.84 0.76 0.84 0.15 0.56 0.16 0.83 0.76 0.88 0.48 0.99 0.81 0.54 0.89 0.59 0.73 0.93 0.77 0.88 0.84 0.80 0.53 0.78 0.87 0.86 0.93 0.72 0.78 0.54 0.73 0.80 0.77 0.30 0.76 0.86 0.76 0.84 0.24 0.74 0.47 0.87 0.34 0.72 Three-Dimensional Control of Color Management Systems M 28 57 85 113 142 170 198 227 255 0.23 0.43 0.27 0.77 0.14 0.23 0.23 0.75 0.45 0.45 0.87 0.12 0.72 0.67 0.87 0.

21 0.28 0.27 0.82 0.83 0.25 0.52 0.13 0.46 0.03 0.86 0.42 0.47 0.85 0.73 0.73 0.89 0.34 0.15 0.23 0.28 0.03 0.03 0.21 0.26 0.41 0.39 0.21 0.21 0.76 0.23 0.84 0.97 0.12 0.21 0.27 0.67 0.06 0.84 0.11 0.85 0.95 0.85 0.52 0.52 0.11 0.85 0.76 0.21 0.89 0.7 (continued) Reﬂectance Spectra for Obtaining Dot Growth Function Reﬂectance Spectra between 400 and 700 nm at 10 nm Interval 0.52 0.26 0.24 0.03 0.25 0.46 0.03 0.73 0.90 0.11 0.06 0.42 0.52 0.90 0.46 0.84 0.84 0.09 0.25 0.90 0.86 0.90 Digital Count 227 255 0.84 0.52 0.81 0.41 0.21 0.07 0.93 1.60 0.02 1.30 0.56 0.85 0.40 0.11 0.18 0.77 0.86 0.03 0.76 0.03 0.76 0.53 0.58 0.14 0.73 0.03 0.25 0.41 0.85 0.52 0.78 0.11 0.95 0.03 0.73 0.41 0.94 0.85 0.21 0.21 0.11 0.11 0.22 0.03 0.86 0.03 0.84 0.41 0.89 0.21 0.11 0.25 0.15 0.43 0.07 K 28 57 85 113 142 170 198 227 255 0.86 0.85 0.11 0.84 0.91 0.84 0.84 0.73 0.83 0.41 0.77 0.03 Control of Color Imaging Systems: Analysis and Design W 0 0.89 0.52 0.11 0.90 0.37 0.83 0.03 0.93 0.11 0.24 0.22 0.77 0.25 0.12 0.40 0.21 0.82 0.72 0.23 0.81 0.01 .54 0.03 0.77 0.11 0.75 0.63 0.40 0.79 0.72 0.41 0.11 0.82 0.41 0.89 0.82 0.40 0.03 0.44 0.03 0.92 0.86 0.83 0.21 0.25 0.86 0.76 0.74 0.72 0.82 0.73 0.52 0.21 0.03 0.41 0.25 0.52 0.03 0.46 0.09 0.76 0.88 0.03 0.73 0.28 0.83 0.76 0.89 0.22 0.25 0.89 0.84 0.52 0.73 0.76 0.42 0.25 0.89 0.77 0.90 0.11 0.21 0.76 0.21 0.76 0.25 0.82 0.76 0.11 0.25 0.06 0.37 0.69 0.52 0.11 0.83 0.41 0.11 0.03 0.83 0.73 0.63 0.83 0.21 0.12 0.15 0.18 0.03 0.16 0.42 0.85 0.11 0.79 0.03 0.25 0.52 0.85 0.25 0.17 0.54 0.83 0.55 0.74 0.90 0.11 0.76 0.342 TABLE 7.05 0.26 0.16 0.78 0.12 0.79 0.53 0.03 0.22 0.83 0.86 0.03 0.03 0.87 0.93 0.83 0.59 0.40 0.21 0.52 0.76 0.25 0.61 0.43 0.86 0.11 0.52 0.76 0.09 0.03 0.06 0.72 0.08 0.88 0.25 0.

6 0.5 0. .7 0.2 0.4 0.9 0.8 0.68. Derivation of these models can be found in many publications.8 0.8 0.3 0.22 A CMY color space is divided into subcubes. These models are.3 0. however.7 0.4 0.5 0.1 0 1 0. and the paper. Kubelka–Munk and Beer–Bouguer models are other models to be considered.8 0. not widely used for halftone printing.9 0.4 0.6 0. The Hoffman–Schmelzer model [66] considers the interaction between the light.3 0.1 0 343 Magenta dot area coverage Cyan dot area coverage 0 50 100 150 200 Digital count 250 300 0 50 100 150 200 Digital count 250 300 Yellow dot area coverage 0 50 100 150 200 Digital count 250 300 Black dot area coverage 0 50 100 150 200 Digital count 250 300 FIGURE 7.2 0.9 0. [38.Three-Dimensional Control of Color Management Systems 1 0.69]. It has been used for the packaging industry with some heuristic corrections [67]. ink.5 0.6 0.21 Dot growth functions plotted as a function of the digital count.7 0.7 0.3 0.5 0.2 0. Refs.4 0. CY Y Grid points in original model (corners of the cube) C Additional grid points in the cellular model (intermediate grid points) MY CM M FIGURE 7.1 0 1 0.6 0.2 0.1 0 1 0.9 0.

Printed colors are selected in critical regions of the color space to maximize the sensitivity to color drift. P(t). The scalar AR model can be used for predicting single channel drift. Consider the measurable output y(n) from the printing system.4 DEVICE DRIFT MODEL A systematic drift in color. at time t0 þ t based on few measurements of printed color patches. is built from the initial forward printer model. and chroma shift or hue shift. we examine a method to predict color drift in digital printers using measurements from a sensor. We describe the general AR model for signal prediction and explain how it can be applied to our speciﬁc problem of printer drift prediction.1 Autoregressive (AR) Model Applied to Printer Drift Prediction Let time t0 be the time at which the printer forward model (map) is initially constructed from measurements. P(t0). The output of the system is given by y(n) at time index n. this would require too many measurements.4. It is easy to build a printer drift model if the measurements for all the patches used for constructing the forward model are available. lightness on the paper. It is assumed that N data samples are available. In this section. The VAR model can predict the color drift speciﬁed in terms of reﬂectance spectra or the L*a*b* on the paper. Thus a drifted printer model.344 Control of Color Imaging Systems: Analysis and Design 7. can be modeled by sampling a few colors and then updating the printer model over time. The function y(n) could be L* or a* or b* or chroma or hue or print density for a single color patch. that is.4. In this model. We use an AR model of order P of a stationary zero-mean process y(n) of the form y(n) ¼ À P X i¼1 ai y(n À i) þ e(n) (7:62) where fai gP are the AR parameters and e(n) is the zero-mean white noise process i¼1 error with variance s2. If performed during run time. the predicted output is given by ^(n) ¼ À y P X i¼1 ai y(n À i) (7:63) The error between the measured output and the predicted output is given by e(n) ¼ y(n) À ^(n) y (7:64) .4. Hence a more preferred method would be to use a few measurements to sample the color drift and to update the entire model. Two methods are discussed: (1) channel independent (scalar) autoregressive (AR) model and (2) channel dependent vector autoregressive (VAR) model. 7. which increase cost and decrease productivity. which may occur due to changes in humidity and temperature. drift in the print density on the photoconductor or on the paper.

we get P P P P X X XX Â Ã ai ryy (i) þ ai ryy (Ài) þ ai aj ryy ( j À i) E e2 (n) ¼ ryy (0) þ i¼1 i¼1 i¼1 j¼1 (7:70) Therefore.Three-Dimensional Control of Color Management Systems 345 To determine the coefﬁcients of the AR model. we have p p X Â Ã Â Ã X E e2 (n) ¼ E y2 (n) þ ai E½ y(n À i)y(n) þ ai E ½ y(n)y(n À i) i¼1 i¼1 p p XX i¼1 j¼1 þ ai aj E½ y(n À i)y(n À j) (7:69) Performing the expectation. the mean-squared error (MSE) between the predicted output and the measured output is minimized.66 can be expanded as "( )( )# P P X X Â 2 Ã ai y(n À i) y(n) þ ai y(n À i) E e (n) ¼ E y(n) þ i¼1 i¼1 (7:67) Expanding the above equation results in " # " # P P X X Â 2 Ã Â 2 Ã ai y(n À i)y(n) þ E y(n) ai y(n À i) E e (n) ¼ E y (n) þ E " þE i¼1 P X i¼1 ai y(n À i) P X j¼1 # aj y(n À j) i¼1 (7:68) Interchanging the expectation and summation. From Equations 7.64.63 and 7. P P P X XX Â Ã ai ryy (i) þ ai aj ryy ( j À i) E e2 (n) ¼ ryy (0) þ 2 i¼1 i¼1 j¼1 (7:71) . the error between the measured output and the predicted output is given by e(n) ¼ y(n) þ P X i¼1 ai y(n À i) (7:65) The MSE can be expressed as the experted value of the square of the error 2( )2 3 P X Â 2 Ã (7:66) ai y(n À i) 5 E e (n) ¼ E4 y(n) þ i¼1 Equation 7.

.. 5 . aP ryy (P) ryy (0) (7:73) Since the output of the system y(n) is real. . 7 6 . so Equation 7. 7 6 . i ¼ 1. 6 . P. 5 4 . 76 .. .. 2. . MSE is given by s2 ¼ ryy (0) þ ryy (1)a1 þ ryy (2)a2 þ Á Á Á þ ryy (P)aP e Combining the above two equations we obtain 2 r (0) ryy (1) ryy (2) yy ryy (0) ryy (1) 6 ryy (1) 6 ryy (1) ryy (0) 6 ryy (2) 6 . aP ryy (P) ryy (0) (7:74) And the variance of the error function e(n). 7 ¼ 6 . ryy (1) 54 aPÀ1 5 0 aP .. 6 4 r (P À 1) yy ÁÁÁ ryy (P) Á Á Á ryy (P À 1) ryy (P) 32 1 3 2 s2 3 e Á Á Á ryy (P À 2) ryy (P À 1) 76 a1 7 6 0 7 7 6 7 76 Á Á Á ryy (P À 3) ryy (P À 2) 76 a2 7 6 0 7 7 76 . ÁÁÁ 2 3 Á Á Á ryy (P À 1) 32 a1 3 ryy (1) ryy (P À 2) 76 a2 7 6 ryy (2) 7 76 7 6 7 76 a3 7 6 7 76 7 ¼ À6 ryy (3) 7 76 . . We minimize the above expression to obtain the optimum values for ai. . 5 4 . . 7 . The autocorrelation and variance of the data can be found using MATLAB functions... As the signal y(n) is real. 7 6 . 5 .346 Control of Color Imaging Systems: Analysis and Design where ryy(i) is the autocorrelation of the output process. 2. 54 . . . . that is.73 can be rewritten as 2 6 6 6 6 6 4 r (P À 2) yy ryy (P À 1) ryy (0) ryy (1) . ryy (0) (7:75) The AR coefﬁcients and the variance of error are determined by solving the above system of linear equations using least-squares regression equations for a set of N number of measured data samples. ryy (1) ryy (2) ryy (0) ryy (1) .. . . 7 . . ryy (1) ryy (2) ryy (0) ryy (1) . its autocorrelation is symmetric. . Applying the orthogonality principle. 7 (7:76) 76 .. autocorrelation function ryy is symmetric (ryy(i) ¼ ryy (Ài)). P (7:72) The above equations can be written in matrix form as 2 6 6 6 6 6 4 r (2 À P) yy ryy (1 À P) ryy (0) ryy (À1) . . we obtain the well-known Yule–Walker equations P X j¼1 ryy ( j À i)aj ¼ Àryy (i) i ¼ 1. y(n). . 6 . . 7 76 . 7 . ÁÁÁ 32 3 2 3 Á Á Á ryy (P À 1) ryy (1) a1 ryy (P À 2) 76 a2 7 6 ryy (2) 7 76 7 6 7 76 a3 7 6 7 76 7 ¼ À6 ryy (3) 7 76 . 7 6 . 54 . 4 5 0 . . . . ryy (i) ¼ ryy (Ài). at ith lag.

The prediction error is assumed to be zero-mean white noise process with unknown covariance matrix S. P are 3 Â 3 matrices which deﬁne the VAR matrix coefﬁcients. Say we have the outputs of the system in L*a*b* and density. After estimating the AR coefﬁcients for the three channels using Equation 7.. y(n) ¼ b*(n) À b*(0) for predicting b*. The model order P can be determined using statistical techniques including the minimization of an order selection criterion [70] (not discussed in this book). measured using a color sensor at time index n corresponding to time t.e.2 Vector Autoregressive Model Applied to Printer Drift Prediction In the vector AR model. This approach ignores the interaction between L*a*b* and d treating them independent of each other. 7. the measured output process is a 3 Â 1 vector random process deﬁned by 2 3 L*(n) À L*(0) y(n) ¼ 4 a*(n) À a*(0) 5 (7:78) b*(n) À b*(0) The predicted output in terms of the P (previously) measured output values is given by the following equation: ^(n) ¼ À y P X i¼1 Ai y(n À i) (7:79) The error signal between measured and predicted outputs is given by e(n) ¼ y(n) À ^(n) ¼ y(n) þ y P X i¼1 Ai y(n À i) (7:80) where Ai for i ¼ 1. d. magenta. The initial time t0 is assumed to be zero. with corresponding time index n ¼ 0. . The error is minimized in the least squares sense (similar to the scalar case). So in the vector AR model. . yellow. . That is ^(n) ¼ c À y P X i¼1 ai y(n À i) (7:77) where c is the initial value of L* for predicting L* and the corresponding initial values for predicting a*. d(n) from the P previous values of the outputs using Equation 7. using the above model (Equation 7. The MSE is given by .4.74) with y(n) ¼ L*(n) À L*(0) for predicting L*.62. independent of the other channels (i. . we can ^ ^ predict the new values L*(n). and density d. â*(n).76. and black). y(n) ¼ a*(n) À a*(0) for predicting a*. y(n) ¼ d(n) À d(0) for predicting print density represented by the symbol d.4. ^ b*(n). We predict the output in each channel (say cyan). or b*.74 or Equation 7. we consider the dependence of one channel on the other channels while predicting the output of the printer.Three-Dimensional Control of Color Management Systems 347 Now we will discuss how this model can be applied to predict the printer drift.

. 0 AP Ryy (P) Ryy (P À 1) Ryy (0) where A0 is a 3 Â 3 matrix with A0(i.81 with respect to matrix Ai. 4 . 54 . . â*(n). We now optimize the cost function given by Equation 7.348 Control of Color Imaging Systems: Analysis and Design Â Ã Â Ã E e2 (n) ¼ E e(n)eT (n) " #" # P P X X T T T E y(n) þ Ai y(n À i) y (n) þ y (n À i)Ai i¼1 i¼1 P X i¼1 ¼ Ryy (0) þ Ai RT (i) þ yy P X i¼1 Ryy (i)AT þ i P P XX i¼1 i¼1 Ai Ryy (j À i)AT j (7:81) where Ryy(i) is the 3 Â 3 correlation matrix of the output of the system at lag i and is given by 2 3 rLL (i) rLa (i) rLb (i) (7:82) Ryy (i) ¼ 4 raL (i) raa (i) rab (i) 5 rbL (i) rba (i) rbb (i) Since y(n) is real. . Comments about the experimental validation of AR drift models are shown next. . . . yy The diagonal elements of the positive deﬁnite matrix Ryy(i) are the autocorrelation of the three components of the color vector L*a*b* and the off diagonal elements are measure of correlation between the three coordinates of the L*a*b* color vector. 7 ¼ 4 5 . . 5 . The result is similar to the Yule–Walker equations for the scalar case and is given by 2 32 3 2 3 Ryy (1) ÁÁÁ Ryy (P) Ryy (0) A0 S 6 Ryy (1) Ryy (0) Á Á Á Ryy (P À 1) 76 A1 7 6 0 7 6 76 7 6 7 (7:83) 6 . Ryy (Ài) ¼ RT (i). and ^ b*(n) from P previous values of the output using the following equation: ^(n) ¼ c À y P X i¼1 Ai y(n À i) (7:85) where c ¼ [L*(0) a*(0) b*(0)]T. 76 . we can predict the new values i¼1 ^ of L*(n). j) ¼ 1 0 is a 3 Â 3 matrix of zero elements S is the 3 Â 3 covariance matrix of the prediction error signal e(n) S is the covariance matrix of the prediction error function and is given by S ¼ Ryy (0) þ Ryy (1)AT þ Ryy (2)AT þ Á Á Á þ Ryy (P)AT 1 2 P (7:84) After estimating the VAR matrix coefﬁcients fAi gP .

7 * DEab 95% 6.4 2. The data were divided into two parts.Three-Dimensional Control of Color Management Systems 60 58 56 54 52 L and D 50 48 46 44 42 40 0 200 400 600 Time index n 800 1000 1200 L* Predicted L* D* 349 Predicted D * FIGURE 7. The * * prediction error statistics. The second data set has 3131 colors in the printer gamut with measurements at 10 different time samples. the data were divided into an estimation and a prediction part. To test the vector AR model.5 5. and 3 is constructed using the ﬁrst 700 samples of the data and the rest of the data is used to verify the model. An AR model of order 1.23 Measured and predicted L on a drifted print engine. Since the data record length is small.7 .8. TABLE 7. A large number of temporal training data set is used for building scalar or VAR drift models. we used interpolation to increase the data set record length.8 2.7 * DEab max 12. The measured and predicted values of L* and D* ¼ 100 À L* for a gray patch are shown in Figure 7. 2. One data set has 33 color patches on the neutral or gray axis with measurements at 1200 different time samples. maximum DEab. and 95th percentile for vector regressive models of order 1. and 3 are tabulated in Table 7.23. average DEab. one part is used to build the drift model and the second part is used for prediction. Again similar to the scalar case.8 Error Statistics for VAR Model VAR Model Order P¼1 P¼2 P¼3 * DEab mean 3.8 10. the second data set is used. * The techniques are quantiﬁed by using DEab ¼ kLabmeasured À Labestimatedk between the actual measured drift data and predicted drift data for the predicted population. We used two sets of printer drift data. 2.9 7.4 4.

as K increases.76 and 7. K ¼ 153 (solid.83). the total gamut volume created by mixing CMY with K decreases (Figure 7. it can be removed from the drift measurements of the printer. . 51. 7.5 GCR SELECTION AND INVERSION In a four-color CMYK printer.24). (c) CMY gamut with K ¼ 0. and K ¼ 255 (bottom). Gray component replacement (GCR) provides a method to substitute black (K) for CMY mixtures in rendering a given color. If the noise model of the color sensor is available.24 (a) CMY gamut with K ¼ 0 (top). which results in an extension of the darker region of the gamut by changing lightness FIGURE 7. (b) CMY gamut with K ¼ 0 (wire). The accuracy of the model depends on the data length as well as sensor noise. middle). and K ¼ 255 (bottom). which can result in a more accurate drift prediction. 102.350 Control of Color Imaging Systems: Analysis and Design The AR and VAR models can predict drift data accurately if the measured data record length is large enough for estimation of the parameters of the models (Equations 7. 204. 255. 153.

25 [71].5x2 from all three-component values of CMY and add black colorant per Equation 7. To avoid this kind of degeneracy and determine a unique combination for four separations. Since the forward map is the representation of a printer with an over-actuated system.1 A SIMPLE GCR FUNCTION A simple GCR function. sky tones. The inverse map can be thought of as cascade of two LUTs: (1) a three-to-three mapping LUT. As the ﬁgure indicates. in control terminology.86. Smoothness and gamut coverage must also be taken into account while rendering pixels with black. M. the four-color forward printer map can be considered as a four-input three-output plant with four input variables (CMYK) used to control three outputs (L*a*b*). CMY. and depends strongly on the physics of the printer. the GCR will be included as a part of the multidimensional LUT. CIELab to CMY and then (2) a three-to-four mapping LUT. For any given node. Experiments are often done with many iterations to get the right amount of K. It helps to reproduce shadows. A potential disadvantage in using high K levels throughout the gamut is the dirty=grainy appearance that can arise in ﬂesh tones. CMY to CMYK. conceptually an RGB node will be transformed ﬁrst to three-color space. An accurate multidimensional proﬁle LUT is intended to provide the inverse of the printer (or inverse of the characterized printer map) for every input color node described as L*a*b* which is a transformation of the node from CIELab into a CMYK color space. Once the tuning is done. CIELab. Numerous techniques are available in the literature [71–80] that use variety of methods to characterize the printer and then compute inverse. Y). there exist many possible combinations of CMYK values that produce a given L*a*b*. 7. print vendors ﬁne-tune the addition of black intelligently either by using complex algorithms or by using carefully designed experiments. (7:86) .5. Among these. For pixels not on the nodes. GCR constraints are required. tetrahedral interpolation is the most commonly used method inside the ICC proﬁle. We will further elaborate on this topic later. The GCR function is described in terms of two components: (1) black addition and (2) under color removal (UCR). then to another three-color space. In such a LUT architecture. Generally.Three-Dimensional Control of Color Management Systems 351 (or darkness) as compared to printers without K separation. and muted tones in images. K ¼ x2 where x ¼ min(C. numerous interpolation schemes are used. These constraints are included in the inverse map (L*a*b* to CMYK LUT). during real-time image processing. is shown in Figure 7. GCR optimization requires a delicate trade-off among these competing requirements. GCR is a critical element in the inversion process since it introduces redundant solutions. Thus. CMYK. This is described next. and other important colors. prior to rendering. ﬁnally followed by a transformation to four-color space. Other beneﬁts of introducing K are toner savings and improved stability. Printer characterization gives an approximation to the forward map (from devicespeciﬁc CMYK space into the device-independent CIELab or spectral space). CMY to CMYK transformation. we ﬁrst subtract 0. gray areas.

355 0. M.25). .3 0.5 0.26 CMYK gamut compared to a CMY to L*a*b* printer gamut with the simple CMY to CMYK GCR (wire: A CMYK to L*a*b* gamut of a four-color printer. Soon it will become clear that this method has the disadvantage of not producing sufﬁciently optimized colors for the entire color gamut since inherently the practice of starting with CMY separations and subsequently adding black leads to loss of gamut. M. However.455 0. CMY ¼ [0.4] color patch (CMY values in 0 to 1 range) is equivalent to CMYK ¼ [0. one gamut produced FIGURE 7.352 Control of Color Imaging Systems: Analysis and Design CMY + Σ – 0. solid: a gamut of a CMY to L*a*b* printer with GCR=UCR of Figure 7.26 shows the comparison between the two gamuts. Figure 7. Y). in Ref.255 0. [81] a max-gamut GCR is developed which starts with CMY separations and still preserves the total gamut volume.5 x2 CMY – x2 CMYK K + = x2 x Min x (a) K + – CMY = 1 x2 2 (b) Min (C.09] patch using the above GCR=UCR algorithm.25 A simple GCR=UCR function (a) block diagram view (b) plot of GCR function with respect to min (C. For example. Y) FIGURE 7.

28 shows a 3-D view of the nodes when constructed with sRGB.g.g. a conversion from RGB to L*a*b* or XYZ is required which is purely an inverse transformation starting from each of the RGB color nodes of the uniformly or nonuniformly sampled RGB grid.g. The CMY values are converted to CMYK using the GCR=UCR algorithms (e. any of the standard RGBs (e.g. ROMMRGB. we apply a suitable transformation in the forward direction from RGB to L*a*b* and map the out-of-gamut colors in L*a*b* to a suitable node on the device boundary (or inside) of the printer model.2 INVERSION OF A THREE-TO-THREE FORWARD MAP The inversion mapping process introduced in the previous section is illustrated in Figure 7. In this section let us consider that the GCR=UCR functions are embedded inside the augmented forward printer map. Since RGB space is not device independent.) are potential candidates for use in the multidimensional proﬁle LUTs. This is an intermediate color space in the ﬁnal ICC destination proﬁle LUT (see Figures 7. Equation 7. where input L*a*b* points are on a three-dimensional (3-D) grid of size N Â N Â N.. Since this is a deterministic forward map. This grid has a dynamic range 0 L* 100. for 17 ) Gamut mapped nodes for out-of-gamut colors as in in-gamut colors are not mapped FIGURE 7. À127 a* 128.27 starting from the intermediate RGB space.g. and genRGB color spaces. Q.. the other gamut produced by a three-color CMY printer (e. Figure 7. ROMMRGB. Figure 7. to produce a new forward map. we start with a uniformly or nonuniformly sampled RGB color LUT in the range 0–255. a CMY to L*a*b* printer map including the GCR) using the GCR function of Figure 7...g.2 GCR/UCR is included with map Q RGB Transformation Labi Gamut mapping Mapped Labi Q–1 Error CMYi Q Labout (Accuracy) Uniformly sampled RGB nodes 3 (e. Now. Next. 7.86).2). To generate the ICC workﬂow. The inverse of this map. to express these RGB nodes (in the reverse direction) in the PCS. for 17 ) Lab nodes 3 (e.1 and 7..25. and À127 b* 128. Once the GCR function is selected using the CMYK to L*a*b* printer model.. CMY to L*a*b*.27 Block diagram illustrating a forward and inverse printer process with GCR in the augmented printer map. it gives unique colors for a unique combination of CMY and does not require any constraints during inversion. Q. an augmented CMY to L*a*b* printer map can be generated by including the GCR function in the system. Inversion of this kind of unconstrained threeinput to three-output forward map is reasonably easy and will be described next. . Q shown symbolically as QÀ1: L*a*b* ! CMY.5. etc. genRGB. a calibrated CMYK to L*a*b* printer map).Three-Dimensional Control of Color Management Systems 353 by a four-color CMYK printer (e.

iteratively clustered interpolation (ICI) algorithm [91]. 3-D root ﬁnding algorithm [92].) also play an important role in the choice of the inversion approach.354 Control of Color Imaging Systems: Analysis and Design FIGURE 7. (1) Shepard’s method is easy to implement but relatively less accurate than others.6. number of complex operations such as multiplications.... The reverse path to PCS space can be found in Figure 7. Although round trip accuracy (Section 7. Hence a more general control-based approach is developed and is described in the next section. conjugate gradient algorithm [93]. . For example. 7. from nodes in the RGB color space to L*a*b*=XYZ values in the PCS) and a correspondence in the forward direction from the RGB nodes to L*a*b* nodes are generated just before constructing the printer inverse (i.4.2.e. and (right: ProPhoto RGB to L*a*b*). (middle: ROMMRGB to L*a*b*). and (3) the ICI algorithm provides by far the best results in terms of accuracy.7) is a good quantitative metric for comparison between various known inversion methods.5. and complexity. Although ICI algorithm is iterative. matrix inversions performed. execution time. This in turn can affect the interpolation accuracy of pixels that are not located on the nodes since all interpolation operations inside the ICC proﬁle workﬂow use linear interpolation such as the tetrahedral interpolation to process images at high speed. a correspondence in the reverse direction (i. computational complexity (i. its use for a nondeterministic system is difﬁcult. It is more suitable for accurately inverting color maps with three or more separations by either working on the printer forward map or by directly iterating on the printer with color sensors.28 Three-dimensional view of node colors when super imposed on a printer gamut (left: sRGB to L*a*b*).1.1 Inverse by Working on the Printer Model Different numerical algorithms can be used to compute the printer inverse map starting from the printer model [71–88]. (2) the MM method has accuracy better than Shepard’s but because of complex operations involved in it. They include Shepard’s algorithm [89].e. Comparison of the inversion performance of some of these methods is described in Section 6. shows the image path in the forward direction when we start from the intermediate RGB space. the moving matrix (MM) algorithm [90].e. The overall proﬁle accuracy depends on the choice of this intermediate color space since the spacing between the nodes in this color space has different effects in different regions of the printer gamut. Various other methods described in Refs. etc. L*a*b* to CMY ). execution time. it has high execution time and the method often fails for nodes near the boundary. [71–88] and [94] can be used in the inversion. Thus.

In Section 8. output of the ICI algorithm). Considering the printer input–output characteristic as linear (which is generally true at the nominal CMY values. However. the system with the integrator in Figure 7. A block diagram shown in Figure 7. see Figure 7.5. a derivation of the linear state space model is shown at the nominal CMY values for a node color whose target L*a*b* values are given. for an individual node color.Three-Dimensional Control of Color Management Systems 355 7.2 Control-Based Inversion We describe the control-based inversion for a linear three-input three-output forward map or a printer. we do not have to constrain the control algorithm when compared to the control approaches applied directly on a nonunique.6.. Since GCR is already embedded inside the augmented printer model (CMY to L*a*b*). In this approach. these values can come from any of the less accurate inversion algorithms described above (e.29 can be expressed in state space form as x(k þ 1) ¼ Ax(k) þ Bu(k) (7:87) where x(k) represents the L*a*b* values from the printer model obtained at iteration k A is an identity matrix Feedback controller Reference. CMYK to L*a*b* printer.30). Closed-loop control algorithm with a gain matrix and an integrator as the . After that. This approach requires the use of nominal CMY values for each node because of the use of linear state space form. Thus.2. which is done by representing printer characterization data in a suitable form such as the state variable form.29 illustrates how inversion operations are executed at a given L*a*b* node.29 controller.g. The problem lies in the design of stable controllers for each of the node colors. The algorithm iterates on the printer model (or the printer) for each in-gamut node by using the L*a*b* node as target. r (L*a*b*) + – e K u Initial CMY values Augmented printer model/ printer Measured L*a*b* x Integrator V + x(k + 1) =Ax(k)+Bu(k) FIGURE 7. We show the use of linear multiple-input multiple-output (MIMO) statefeedback controllers [95] to update the ﬁnal CMY values that will further reﬁne the inversion errors. we design the feedback controller for this system. a feedback control algorithm can be used at each node to accurately convert the in-gamut L*a*b* nodes to CMY color space. we ﬁrst develop a state space model for the CMY to L*a*b* printer model (or the printer).

where e(k) is the error between the target L*a*b* and the model L*a*b* at iteration k. Thus. u(k) ¼ ÀKe(k). the eigenvalues of the closed-loop system should remain inside the unit circle of the Z-domain even when the printer drifts. the pole values .1) in the Z-domain. This is achieved by assigning pole values within the range [0. In poleplacement design (Chapter 6). the gain matrix K is derived based on the pole values speciﬁed such that the closed-loop system model shown in Figure 7. For the loop to be stable.30 M and Y.356 100 80 60 40 20 0 40 20 b* 0 Control of Color Imaging Systems: Analysis and Design 100 50 o a* 0 L* o –50 –100 50 Cyan (%) 100 0 50 Cyan (%) 100 M = 50%. [95]) u(k) is the control law applied to the input of the printer The Jacobian matrix B is different for each node color and computed as follows: @L* 6 @C 6 6 @a* B¼6 6 @C 6 4 @b* @C 2 3 @L* @L* @M @Y 7 7 @a* @a* 7 7 @M @Y 7 7 @b* @b* 5 @M @Y (7:88) The control law is designed using MIMO state-feedback controllers (based on linear quadratic regulator [LQR] or pole-placement as shown in Chapter 6 or any other multivariable digital control techniques [96–100]). Hence. Y = 50% o M = 10%. Diagram representing L*a*b* values when C separation is varied at constant B is the Jacobian matrix computed around the initial CMY value (or measured using techniques outlined in Ref. Y = 50% o Nominal point –20 –40 0 50 Cyan (%) 100 FIGURE 7.29 is stable.

e. CMY).37 0. the DE errors shown in Figure 7.18 0.Three-Dimensional Control of Color Management Systems ΔE wrt node L*a*b* in CIELab space 70 60 50 40 30 20 10 0 0 2 4 6 8 10 12 14 16 18 20 Iteration No. which can result in the choice of wrong inverse (i. Nominal CMY values can also be calculated using other algorithms (e.77 have to be carefully tuned to avoid oscillations during iteration (or instability) and potential actuator saturation for near-boundary colors. a unique best actuator algorithm shown in the diagrams of Figure 7.e. This plot clearly indicates that the convergence error (shown in * Table 7.41 0.86 0. FIGURE 7.58 Mean 2.31 for node colors at the gamut boundary.01 * DEab 95% 9.00 Max 62. During iterations. Assume that there are N0 colors that will be adjusted by the algorithm. which is a signiﬁcant improvement when compared with other inversion algorithms used in the industry.2. The best CMY selection algorithm starts from the second block in Figure 7.3.9 for both DEab and DE2000 ) is near zero for all the nodes at the end of the iterations. meaning DE descent with respect to iterations can oscillate.g.67 0.3).00 Max 60. ICI from Section 6. which will iterate Nt times and all iteration history is stored. 70 60 50 40 30 20 10 0 0 2 4 6 357 ΔE wrt node L*a*b* in 2000 space 8 10 12 14 16 18 20 Iteration No. Since iterations contain the history of oscillations. node colors near the boundary may reach saturation (i.31 Convergence plot for 5192 in-gamut nodes using nominal values calculated from MM algorithm.32 can be used to select proper CMY values that give the best inverse for the node colors at the gamut boundary.9 Accuracy Statistics for In-Gamut Nodes before the Start and End of Iterations DE2000 Mean Start of iteration End of iteration 1.00 95% 6.18 0.. Figure 7.. their CMY values may go to 0 or 255).32a where the algorithm picks the .4. That is.. may go higher in a more recent iteration than the previous one. TABLE 7.31 shows the convergence plot for all the in-gamut nodes when the nominal CMY values are calculated using the MM algorithm mentioned in Section 6. This can lead to limit cycles during control iterations.

Nt End Nt+ 1 Nt+ 2 Set used for best CMY Set used for best CMY Set used for best CMY (b) FIGURE 7. (b) Best CMY selection progression for control-based algorithm shown for Nt iterations and beyond.32 (a) Best CMY selection ﬂow diagram for node colors shown for Nt iterations and Nc colors. the algorithm continues the process for the next color in the list. ﬁrst color for the analysis. 2. and many other control system-related shortcomings are overcome by using a multiplicity of gain matrices per node color as opposed to single gain matrix. These.358 Control of Color Imaging Systems: Analysis and Design 1 Run Nt iterations of control algorithm 2 Initialize c = 1 (Start with first color) 3 Select (CMY )c. The third block seeks CMY values for each color c that generated the minimum DE value across all Nt iterations. The fourth block checks whether all colors considered in the algorithm have already been analyzed. the selection algorithm stops and the best CMY values selected will be used for the respective node colors. . Thus the best actuator algorithm selects the best CMY that leads to minimum DE convergence by assessing DE results across all the iterations. Linear state-feedback design also has other shortcomings such as not reaching zero steady state error for colors near the boundary. If so. i Nt 4 c > Nc N c = c+1 Y (a) 1.i* i *= arg mini {ΔE 2000c. The selected CMY values are stored. An example is shown in Problem 7.i}.7. …. Otherwise.

most of the neutrals are produced using CMY separations and no black.7. process the error between the measured L*a*b* values and the target L*a*b* values.8 Generate a GCR constrained ICC destination proﬁle with a 3-to-3 control-based inversion using the GCR function of Figure 7. Compare the effective volume with respect to the gamut volume obtained from the characterization data.6. Improvement of accuracy in performance for boundary colors can also be achieved by scheduling multiplicity of gain matrices for every node color using MIMO model-predictive-control methods [100].2. Find nominal CMY values for the node colors of interest using ICI or other inversion algorithms on a coarse printer model. a. such as contours for dark colors if left unattended. The iteration process contains the following steps. an L* linearity plot is shown. The L*a*b* values of the color patches on the printed test image are measured.. The CMYK response of the ICC proﬁle for RGB neutral sweep is shown in Figure 7. Evaluate the visual response of the ICC proﬁle for a test image. gain matrix and the integrator). Continue with steps a through c for a few iteration cycles. Also a sharp transition in the cyan separation for dark colors can cause problems. the historical methodology of building a printer model and then inverting that model for developing multidimensional LUTs may not be necessary. Example 7. This is good for high quality rendering. [96–100].g.25 for a test printer. Clearly. In Figure 7.34.Three-Dimensional Control of Color Management Systems 359 7. The best actuator selection method described in the previous subsection may be required for node colors near the gamut boundary.8 and the GCR of Figure 7.3 Inverse by Iterating Directly on the Printer If we can directly iterate on the printer for any node color using control algorithms. The corresponding CMYK values are calculated for these nominal CMY values. The steps involved in evaluating the ICC proﬁles are covered in detail in Section 7.25. c.5. This could cause image quality defects. A coarse printer model is a less accurate representation of the printer. Evaluate the quantitative round trip accuracy and gamut volume. Clearly. using the CMY to CMYK GCR function. Use a suitable gamut-mapping strategy described in Section 7. b. The gain matrix can be found using pole-placement or LQR methods a priori at the nominal CMY values or using any of the control approaches shown in Refs. we see problems with dark colors (once again) that is a problem at . d. Using the controller shown in Figure 7. The rapid rise of magenta and yellow separations between R ¼ G ¼ B ¼ 0 to about 50 digital counts is caused by mapping out-of-gamut colors to the printer gamut.29 (e. Create a test image containing the color patches of the determined CMYK values and print. SOLUTION We constructed a 333 ICC proﬁle using the steps shown in Section 7.33.

. 2200 2000 1800 1600 1400 1200 1000 800 600 400 200 0 * Max ΔEab Avg ΔE* ab 53. right: controlbased inversion).8953 8.2994 0 2 4 6 8 10 12 14 16 2200 2000 1800 1600 1400 1200 1000 800 600 400 200 0 Frequency of occurrence Frequency of occurrence * Max ΔEab * Avg ΔEab 25.9398 0 2 4 6 8 10 12 14 16 * Color difference ΔEab * Color difference ΔEab * FIGURE 7.2948 3. right: controlbased inversion). right: control-based inversion). 100 90 80 70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 80 90 100 sRGB input L* along a* = b* = 0 100 90 80 70 60 50 40 30 20 10 0 Print output L* along a*=b*=0 Print output L* along a*=b*=0 0 10 20 30 40 50 60 70 80 90 100 sRGB input L* along a* = b* = 0 FIGURE 7.34 Input L* vs.33 GCR response to neutral RGB sweep (left: with MM inversion.35 Histogram of roundtrip accuracy shown in terms of DEab (left: with MM inversion.360 CMYK response to neutral 8 bit input 250 200 150 100 50 0 Control of Color Imaging Systems: Analysis and Design CMYK response to neutral 8 bit input 250 200 150 100 50 0 Cyan Magenta Yellow Black Cyan Magenta Yellow Black 0 50 100 150 200 250 <----Black White---> R = G = B 8 bit input (a*= b*= 0) 0 50 100 150 200 250 <----Black White----> R = G = B 8 bit input (a* = b* = 0) FIGURE 7. output L* response (left: with MM inversion.

Clearly. Figures 7. blocking.36). This is largely associated with the multidimensional proﬁle LUT with the 3-to-3 inversion and GCR. Most round trip colors are not on the grid. etc. Also.36 (a) Vector plots between colors produced with A2B1 (forward) and colors produced with combined maps (B2A1 (inverse) and A2B1 (forward)) with MM-based inversion (left: view from the bottom. * the bottom of the gamut for pixels outside of it.40 show the visual effects of inversion algorithms on various images as compared to the original image. most inaccuracies are near the region below the CMY gamut (see Figure 7.) can be associated to inversion algorithms.35. all other things being equal.Three-Dimensional Control of Color Management Systems 361 (a) (b) FIGURE 7. . A DEab histogram for in-gamut round trip colors is shown in Figure 7. objectionable color contents (shifts. contours. (b) Vector plots between colors produced with A2B1 (forward) and colors produced with combined maps (B2A1 (inverse) and A2B1 (forward)) with control-based inversion (left: view from the bottom. the effective gamut utilization with this kind of GCR was found to be around 84%. Although nodes have converged to nearly zero DE. These errors can be associated to GCR methods used in this process.37 through 7. The proﬁle inaccuracy is largely due to the inaccuracy caused by interpolation. right: view from the side). right: view from the side). which means a loss of 16% printable volume.

In another method. bottom: simulated with MM inversion. In Ref. One method is black addition in which black (K) is calculated as a function of a scaled inverse of L*. a three-input four-output transform. subject to certain constraints.) Lena image rendered with ICC proﬁles from Example 7. in . several methods for determining the black (K) component are reviewed. middle: simulated with control-based inversion).362 Control of Color Imaging Systems: Analysis and Design FIGURE 7. is used to calculate the black component. The constraints placed on the transform include a requirement for the sum of the color component values at a node to be less than a threshold.8 (top: original. 7. For example. such as C.5. black (K) is calculated as a function of the minimum value of the other color components. and Y for the CMY color space.3 BRIEF REVIEW OF GCR METHODS The GCR methods available in the literature can be divided into two categories: (a) transformations with ﬁxed GCRs and (b) transformations with ﬂexible GCRs. In a third method. Fixed GCRs are most commonly used in multidimensional LUTs. Flexible GCRs allow the user to have control over the gray component across all levels of GCRs. M. [71].37 (See color insert following page 428.

38 (See color insert following page 428. middle: simulated with MM inversion.Three-Dimensional Control of Color Management Systems 363 FIGURE 7.) Image showing the simulation of Daffodil plant (left: original image. top: MM inversion.) Hairs image rendered with ICC proﬁles from Example 7.39 (See color insert following page 428. right: simulated with control-based inversion). FIGURE 7.8 (left: original. . bottom: control-based inversion).

and in Ref..e. and consistency (i. Although the Pareto-optimal approach. [87]. [104] generate LUTs to minimize metameric effects. and (3) determining the other color components.) Image showing the rendering of color sweeps with ICC proﬁle from Example 7. In this method. a ﬂexible method for estimating the black component comprises (1) determining maximum black component. CMYK color space.4 GCR CONSTRAINED 4-TO-3 INVERSE The CMY gamut (referred to as the gamut with K not equal to zero and with CMY to CMYK GCR) is generally limited by the type of GCR function used as compared to the full CMYK gamut. But this method has a disadvantage of not producing sufﬁciently optimized colors for the entire color gamut. In another method. the UCR=GCR strategy is used to characterize the moiré as a function of the color components and to select the optimized output color components when the moiré function is minimized. right: control-based inversion). Various other GCR control strategies are shown in [84–86]. one of the more ﬂexible GCR methods.8 (left: original image. Ref. provides good control of multiple GCR schemes with full CMYK gamut utilization and a controlled substitution of black into a single methodology. 7. middle: MM inversion. [103] to generate ICC proﬁle. a colorimetric accuracy match) is introduced through carefully selected optimization parameters that attempt to use unique CMYK solutions during the inversion process and yet preserve arbitrary GCR ﬂexibility within a printer’s reproducible limits.40 (See color insert following page 428. C þ M þ Y þ K would be constrained to be less than a threshold. (2) adjusting the black component amounts based on chroma. [82. In Refs. A second constraint is to maintain K between certain minimum and maximum values. it can lead to loss of accuracy during inversion. Other inversion methods are shown in Ref. the designer should be cognizant of .364 Control of Color Imaging Systems: Analysis and Design FIGURE 7.83]. Due to this limitation.5. another UCR=GCR strategy is proposed in which the optimization is done to reduce moiré.102] ICC-based proﬁle LUTs are built to convert images to CMYK. In the Paretooptimal approach [101.

Figure 7. No augmented printer or printer model is used.5. L*a*b* are generated by applying suitable transformations on a uniformly= nonuniformly sampled RGB color grid (e. genRGB). Printer=printer model Jacobian (Equation 7.g. In addition vectors {x.41 shows the block diagram of the 4-to-3 control-based inversion. This makes the interactions between gamut mapping and the GCRs even more complex. b. u. A 4-to-3 control-based. a GCR strategy is deﬁned in the CMY to CMYK transformation. whereas.. 7. Feedback controller (Reference r) L*a*b* + – e K u Integrator Initial CMYK values + V Printer model x(k) x (k +1)=Ax(k) +Bu(k) Measured L*a*b* FIGURE 7. As in previous method (Figure 7. This allows the use of full CMYK gamut when creating the inverse. which is clearly very similar to Figure 7.1 A 4-to-3 Control-Based Inversion The 4-to-3 control-based inversion is applicable to a four-color printer.Three-Dimensional Control of Color Management Systems 365 three regions in the L*a*b* color space for color reproduction: (1) the out-of-gamut region.29. to use the same gamut volume for minimum or maximum black. The key differences between 3-to-3 and 4-to-3 inversion are a. Hence a preferred approach would be to completely eliminate the loss of gamut due to GCRs. in the 3-to-3 inversion. ‘‘Initial (or nominal) CMYK values’’ contain the GCR strategy for in-gamut node colors. When the available gamut is limited by the CMY gamut. c. the out-ofgamut colors have to be mapped to the boundary of the CMY gamut as opposed to the CMYK gamut.41 Closed-loop control algorithm with a gain matrix and the integrator as controller for a 4-to-3 control-based inversion. and (3) the region outside the CMY gamut but within the CMYK gamut. that is. V} and matrices {A. constrained inversion generated by (a) iterating on the printer model or (b) by directly iterating on the printer can offer full CMYK gamut for any GCR curve and high accuracy inversion to node colors inside the gamut. .89) and gain matrices contain the sensitivity of output color to black separation.4. C} in the state variable model and the feedback contain terms associated with black separation. (2) the CMY gamut.29) the node colors in the device-independent color space. B.

best actuator algorithm (similar to the one described in Figure 7. The initial estimated values for black (K) are then used to produce values for other separations. The pole-placement algorithm.2. Figure 7.4. ICC proﬁle) is generated based on the ﬁnal L*a*b* to CMYK LUT. Gain scheduling algorithms based on model-predictive control technology [100] are other methods suitable for use in the control algorithm.32) would be required for node colors near the boundary to select the best CMYK values during the iteration run. in the initial CMYK LUT of Figure 7.42 shows two graphs of the function generated by Equations 7.3) or LQR methods (see Section 5.42.366 Control of Color Imaging Systems: Analysis and Design @L* 6 @C 6 6 @a* B¼6 6 @C 6 4 @b* @C 2 @L* @M @a* @M @b* @M @L* @Y @a* @Y @b* @Y 3 @L* @K 7 7 @a* 7 7 @K 7 7 @b* 5 @K (7:89) A ﬁnal multidimensional proﬁle LUT (i.106] (see Section 5. selected based on the color gamut. h. 7. Furthermore. The graph on the right contains K values (vertical axis) as a . which is the sum of the initial (or nominal) LUT containing the GCR and the newly found correction DCMYK (vector V in Figure 7.5. M.91. The Jacobian matrix at the nominal CMYK values for each node color is used to compute the gain matrix.3.e. The inversion approach uses a 3-to-4 control-based algorithm with a MIMO gain matrix.2 K-Restricted GCR The function used to generate the black (K) component can be a parametric multidimensional function known as the K-function or K-restricted GCR function.90 and 7. The K-function produces the ﬁrst (initial) estimated values for the black component based on the L*a*b* values of the node colors. The graph shown on the left side is the L* response (vertical axis) with respect to K values (horizontal axis). C. place() [105. and Y. The multidimensional K-function may be deﬁned by the following equation: K¼ where 0 B U ¼ 2B @ e Àb U h eÀa a* þb* À 1 2 2 Á =L* 2 1 À Uh þ1 2 (7:90) 1 1 L * L0 ÀL* þ1 2 C À 0:5C A (7:91) and L0..41) through iterations on the CMYK printer or the CMYK printer model. b. and a are predetermined values.2) are candidate algorithms used for computing the gain matrix.

100 0 90 50 80 70 100 60 L0 L* 50 150 30 200 Black digital acount 40 α η L*=10 L*=20 L*=30 L*=40 L*=50 L*=60 L*=70 L*=80 L*=90 L*=100 20 10 0 250 –100 –80–60 –40 –20 0 20 40 60 80 100 0 250 200 150 100 50 Three-Dimensional Control of Color Management Systems Black digital count a*. (b) color gamut (center). 367 . and (c) K digital count with respect to a* and b*. and h ¼ 0.42 A graph of K-restricted multidimensional GCR function generated with a ¼ 0. (a) L* with respect to K digital count (left). b* FIGURE 7. b ¼ 1.01.5.5.

01. while a small value of a means that K is mostly determined by luminosity. h.5. b ¼ 1. These curves show that the K values are weakly dependent on either the a* or the b* values. A large value of a gives a strong dependence of K on a* and b*. Figure 7. and h ¼ 0. which is the place where most colors require adjustment to obtain pleasing appearance on the paper. Larger value of h means the 300 L0 = 102 L0 = 110 L0 = 160 250 200 Black amount 150 100 50 0 0 10 20 30 40 50 L* 60 70 80 90 100 FIGURE 7. The parameter h determines the shape of the K-function curve at small L* values (see left side of the graph in Figure 7. K ! 0 when L* ! L0.368 Control of Color Imaging Systems: Analysis and Design function of a* or b* (horizontal axis) and using a* ¼ 0 (against b* as the horizontal axis) or b* ¼ 0 (against a* as the horizontal axis) for different values of L*. Since L0 is selected to be around 100 (top of the gamut). For a digital production printer gamut.5.5.5 shown for three different values of L0. b.5.43 A neutral response of K-restricted GCR functions with a ¼ 0. and h ¼ 0. 2. only a small amount of black will be used near the white. and h are a ¼ 0. The color gamut with L* (vertical axis) and a* (horizontal axis) is shown in the center sandwiched between these graphs.42).5. b ¼ 1. b.01. . typical values used for a. and h ¼ 0. b ¼ 1. and a are parameters of the multidimensional K-function which satisfy following properties: 1. L0. Parameter L0 affects the middle part of the curve near the neutral region (shown by double arrow).01. 3. The function is symmetrical with respect to a* and b*. K ! 1 when L* ! 0 so that maximum value of K is used at the dark end of the color space. and K increases when the color is away from the neutral zone.43 shows K as a function of L* for a* ¼ 0 and b* ¼ 0 for three different values of L0 with a ¼ 0.

d. the simultaneous perturbation stochastic approximation (SPSA) method [107] can be used to determine a local region where the parameters could be trained using boundary values. . This data can be obtained from the printer model (e. c. Repeat steps (b) through (f) until the CMYK values are found for all the in-gamut node colors.. Find all DE2000 values between node L*a*b* value and the L*a*b* values from the characterization LUT of step (a).g.5 using the ﬁltered K-restricted proﬁle as the source. b.9 uses these steps to create a K-constrained ICC proﬁle. Run a 4-to-3 iterative control algorithm of Figure 7. If experimental characterization input–output data is available. Pick few colors that are closest to K value in the K-function.91. h.91). g. This gives the K-restricted GCR constrained CMYK value for the node color. Compute the Jacobian and gain matrices for each of the colors picked in step (d). Record their corresponding CMYK values.Three-Dimensional Control of Color Management Systems 369 K-function curve is more rounded at small L*. but these inaccuracies can be minimized by running the steps described in Section 7. Filtering may make the in-gamut CMYK values inaccurate. L0) to be optimized and the search space is huge. To set the parameters of Equations 7. f. then use numerical interpolation techniques described in Chapter 6 in place of the printer model. Use the CMYK values from step (g) to complete the ICC proﬁle.41 on a printer model. Complete procedure that is used to create the K-restricted (GCR constrained) L*a*b* to CMYK LUT is shown below. h. which would involve applying multidimensional ﬁltering techniques [108] (Chapter 6) to soften the kinks. the information at the boundary of the color gamut can be used since there is unique correspondence between L*a*b* to CMYK at that location.90 and 7. e. Create a uniformly=nonuniformly sampled high density CMYK to L*a*b* LUT. Example 7. Compare the DE2000 numbers and select CMYK values with the smallest DE2000 numbers.5. a 164 CMYK to L*a*b* LUT). Compare K values between steps (b) and (c) for each of the colors picked in step (c).90 and 7. If the ICC proﬁle has undesirable kinks in CMYK response between neighboring nodes. then another preconditioning step is required. The following steps are described for one in-gamut node color: a. Because there are four parameters (a. Calculate the K value for the node color (L*a*b* data) using the K-function (Equations 7. The parameter b determines the shape of the K-function in the middle range of L*. These data can also be obtained by printing a set of patches and measuring L*a*b* values with a sensor. Rank order the colors with respect to DE2000 values and pick all colors with DE2000 < 6 (or any reasonable threshold which is set depending on the desired accuracy). There could be multiple CMYK values for any given node color. b.

K ¼ 255).91. Solid curves represent the responses before ﬁnal ﬁltering. the differences between solid and dashed curves are negligible as compared with similar ﬁgures in Figure 7. K-constrained control-based algorithm. In this example. there is signiﬁcant loss in the round trip accuracy.2 to understand more about the corner plots). dark red (C ¼ 0. the proﬁle LUT is. Without the control-based inversion on the post-ﬁltered CMYK values. To avoid the formulation jumps (i.6) combined with additional iterative control procedure on ﬁltered data may be necessary to soften the kinks in CMYK formulations. As shown in this example.10 shows the round trip accuracy at the end of each preconditioning step. Clearly. . Table 7. It is to be noted that each precondition step includes ﬁltering and 4-to-3 control-based inversion (Figure 7.7.41) carried out using post-ﬁltered CMYK values as the initial CMYK LUT in the inversion process. We ﬁlter the LUT by applying the multidimensional smoothing algorithm and reﬁne the accuracy by applying a 4-to-3 control-based inversion on the printer model using post-ﬁltered CMYK values as the starting points for the inversion. Also.9 Construct a K-restricted GCR for a uniformly sampled printer input–output data using the analytic functions of Equations 7. Filtering is applied to all nodes (i. Y ¼ 0. In Figure 7. Figures 7. with limited node size (333). we see an effective gamut utilization of greater than 94%. in the right direction. These plots are drawn from white (C ¼ M ¼ Y ¼ K ¼ 0) to the dark black (C ¼ 0.44 shows the CMYK response curves (also called corner plots) for a ﬁrst inverted LUT produced with a 4-to-3.e. The x-axis is extended to reach out-of-gamut colors until it touches the sides of the cube formed by the device-independent space. Except for a nonsmooth CMYK response. Y ¼ 0. non-smooth CMYK response). The inversion algorithm tries to follow the general trend. the K response shown by dashed dot curve in Figure 7. we obtain a ﬁnal inverted LUT to produce smooth CMYK response (see response curves in Figure 7. The CMYK formulations between adjacent nodes have to be smooth.90 and 7. After performing this step a few times.370 Control of Color Imaging Systems: Analysis and Design Example 7.47 and 7. M ¼ Y ¼ K ¼ 255). M ¼ 0. step (c) can be started with previously inverted LUTs using methods such as the MM.e.44. multidimensional ﬁltering (Section 6.91 and a 4-to-3 control-based inversion approach. They are obtained after the ﬁnal ﬁltering step.45) due to nonsmooth CMYK formulations caused by the inversion algorithm. Show the advantages with respect to gamut and image quality (IQ)..h) is calculated from Equations 7.f. Alternatively. Y ¼ K ¼ 255) and dark blue (C ¼ M ¼ 255. M ¼ 0.44 (d. in general.90 and 7.46 shown just before and after the sixth preconditioning step). we apply the preconditioning steps mentioned in step (h) above and generate the second inverted LUT. dark green (C ¼ 255.46. K ¼ 255) corners of the gamut (see Section 7. Values 0–100 on the x-axis represent points starting from white with 100 on the corner. in-gamut and out-of-gamut nodes) in the multidimensional proﬁle.48 show the image quality improvement due to preconditioning steps.. SOLUTION Figure 7. the CMYK response shown by dashed curves for both in-gamut colors and out-of-gamut colors is not smooth. The dashed curves are smooth. Solid curves represent CMYK response after one post ﬁltering step. This LUT produces contours in some images (see the ‘‘neck’’ region shown in Figure 7. This is because the starting CMYK values obtained in step (c) contain local minima (depending on the resolution of the CMYK grid).

44 Inversion response curves for K-restricted GCR (without preconditioning steps). (continued ) . (a) C and M response toward dark black corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)). (b) Y and K response toward black corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)).250 200 150 100 50 0 0 20 40 60 80 From 0 to gamut corner Corners = 0 0 0 255 Y-filt Y-unfilt Corners = 0 0 0 255 250 Corners = 0 0 0 255 C-filt C-unfilt 200 150 CMYK values 50 0 100 250 K-filt K-ref K-unfilt 0 20 CMYK values 100 40 60 80 From 0 to gamut corner 100 250 200 Corners = 0 0 0 255 M-filt M-unfilt 200 150 150 100 CMYK values Three-Dimensional Control of Color Management Systems CMYK values 100 50 50 0 0 100 0 20 (a) 40 60 80 From 0 to gamut corner 0 20 (b) 40 60 80 From 0 to gamut corner 100 371 FIGURE 7.

.44 (continued) (c) C and M response toward dark red corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)). (d) Y and K response toward dark red corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)).372 Corners = 0 255 255 255 250 200 150 100 50 0 0 Corners = 0 255 255 255 250 200 150 100 50 0 K-filt K-ref K-unfilt 20 40 60 80 From 0 to gamut corner Y-filt Y-unfilt Corners = 0 255 255 255 250 200 C-filt C-unfilt CMYK values 150 50 0 100 CMYK values 100 0 20 40 60 80 From 0 to gamut corner 100 250 Corners = 0 255 255 255 200 M-filt M-unfilt 150 CMYK values 50 0 100 CMYK values 0 100 0 20 Control of Color Imaging Systems: Analysis and Design (c) 40 60 80 From 0 to gamut corner 20 40 60 80 100 (d) From 0 to gamut corner FIGURE 7.

(continued ) .Corners = 255 0 255 255 250 200 150 100 50 Y-filt Y-unfilt Corners = 255 0 255 255 250 200 C-filt C-unfilt 150 CMYK values 50 0 0 Corners = 255 0 255 255 250 200 150 100 50 K-filt K-ref K-unfilt 20 40 60 80 From 0 to gamut corner 0 100 CMYK values 100 0 20 40 60 80 From 0 to gamut corner 100 Corners = 255 0 255 255 250 200 M-filt M-unfilt 150 Three-Dimensional Control of Color Management Systems CMYK values 50 0 100 CMYK values 0 0 100 0 20 (e) 40 60 80 From 0 to gamut corner 20 (f) 40 60 80 From 0 to gamut corner 100 373 FIGURE 7. (f ) Y and K response toward dark green corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)).44 (continued) (e) C and M response toward dark green corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)).

.44 (continued) (g) C and M response toward dark blue corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)).374 Corners = 255 255 0 255 250 200 150 100 50 0 0 Corners = 255 255 0 255 250 200 150 100 50 0 90 100 K-filt K-ref K-unfilt Y-filt Y-unfilt Corners = 255 255 0 255 250 200 C-filt C-unfilt 150 CMYK values 50 0 90 100 0 10 20 30 40 50 60 70 80 From 0 to gamut corner CMYK values 100 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner Corners = 255 255 0 255 250 200 M-filt M-unfilt 150 CMYK values 50 CMYK values 0 100 0 0 10 20 10 20 30 40 50 60 70 80 90 100 Control of Color Imaging Systems: Analysis and Design (g) 30 40 50 60 70 80 From 0 to gamut corner (h) From 0 to gamut corner FIGURE 7. (h) Y and K response toward dark blue corner (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)).

8 with 3-to-3 control-based inversion. We can incorporate different ﬁxed GCR=UCR approaches as starting LUTs as future upgrades since each of these LUTs can be ﬁne tuned intelligently using carefully designed experiments in the factory. .Three-Dimensional Control of Color Management Systems 375 FIGURE 7.41 on a drifted print engine in the ﬁeld. For example.45 (See color insert following page 428. a faster way to build the proﬁles is to run the procedure in two groups. middle: simulated using ICC proﬁle from Example 7. there are other beneﬁts in carrying the GCR-constrained initial CMYK and gain matrix values to the ﬁeld.44. For this kind of architecture. The above procedure may take more computing to complete all the steps. Apart from the speed advantage. The ﬁrst group can be used to build the initial CMYK LUT and the gain matrix in the factory on a nominal print engine for all the node colors (includes in-gamut and out-of-gamut node colors). Hence. These initial CMYK and gain matrices can be carried as source=starting LUTs for the control algorithm. (left: original. right: simulated with K-restricted 4-to-3 control-based inversion without preconditioning [step (h)]). The second group can be used for running the iterative control algorithm of Figure 7. especially due to large time spent in step (c). This can be done by extrapolating the printer model for the nominal state or through extrapolation of the initial CMYK values for the in-gamut node colors produced on a printer under nominal state.) Hair image rendered with K-restricted ICC proﬁle whose inversion response is shown in Figure 7. it is important to note that the initial CMYK and gain matrix values need to be found for colors that the printer cannot render in its nominal state but are required to cover the drifted printer gamut. a maximum=medium=minimum black solution can be easily created by adjusting the parameters of the GCR equations and still achieve full CMYK gamut utilization for images by maximizing the accuracy for all the in-gamut node colors. These and other extensions to multiple GCR=UCR combinations can be comprehended within the single initial LUT structure.

376 Corners = 0 0 0 255 250 200 150 Corners = 0 255 255 255 250 200 150 100 50 0 0 20 K-unfilt C-unfilt M-unfilt Y-unfilt K-ref K-filt C-filt M-filt Y-filt CMYK values K-unfilt C-unfilt M-unfilt Y-unfilt K-ref K-filt C-filt M-filt Y-filt CMYK values 100 50 0 100 0 20 (a) (b) 250 200 150 100 50 0 40 60 80 From 0 to gamut corner 40 60 80 From 0 to gamut corner Corners = 255 255 0 255 100 Corners = 255 0 255 255 250 200 150 K-unfilt C-unfilt M-unfilt Y-unfilt K-ref K-filt C-filt M-filt Y-filt CMYK values K-unfilt C-unfilt M-unfilt Y-unfilt K-ref K-filt C-filt M-filt Y-filt CMYK values 100 50 0 100 (c) 0 20 40 60 80 From 0 to gamut corner (d) 0 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner Control of Color Imaging Systems: Analysis and Design FIGURE 7. .46 Response curves for K-restricted GCR (after preconditioning iterations) to different gamut corners from white (C ¼ M ¼ Y ¼ K ¼ 0)).

or control-based). ICI. top right: processed image with ICC proﬁle from K-restricted GCR after one step [contours are visible].Three-Dimensional Control of Color Management Systems 377 FIGURE 7.) Image showing improvements to contours with preconditioning steps for K-restricted GCR (top left: original image. Use of partitioned gamuts provides the following advantages: 1. This approach reduces the dimensionality of the four-color process to three-color groups.47 (See color insert following page 428. the overall color printer gamut can be represented as a composite of the gamut subclasses. wherein each gamut subclass is comprised of a subset of printer color separations [109].4. and MYK ! L*a*b* sub-gamuts that cover the whole printer gamut. 2. 7.3 Tricolor GCR In some printers. bottom: same after preconditioning steps [contours are removed]). .5. Nodes are assigned to one of the gamut subclasses for efﬁciently calculating the CMYK color separation values using one of the inversion techniques (MM. Improved toner usage for high area coverage printing by identifying the most toner efﬁcient CMYK values for in-gamut colors. A three-gamut class method classiﬁes the colors into CYK ! L*a*b*. Improved node color accuracy with one of the separations always held to zero. CMK ! L*a*b*.

49 shows the CMK.48 (See color insert following page 428. Provides more room for the controller on the actuators when the controlbased inversion is used.06 2. C. there exists . M.13 1.21 Max 23. MYK. 5. is always equal to zero for any given color.77 29.98 7.06 6.63 29. right: same after preconditioning steps [contours are removed]).18 6.72 5.23 7. middle: processed image with ICC proﬁle from K-restricted GCR after one step [contours are visible].01 2. Since one of the toners.08 1.52 6.75 1. Figure 7. In a typical four-color printer. TABLE 7.53 Max 12.82 27. 4.39 4.08 * DEab 95% 3. or Y.14 6.29 4.54 5.) Image showing the rendering of color sweeps (left: original image.16 1.10 Round Trip Error with Respect to Number of Iterations Used in the Preconditioning Method with Filtering Step Included in the K-Restricted GCR Algorithm Preconditioning Steps None 1 2 3 4 5 6 DE2000 Mean 0.90 1.82 3.47 3.79 6.20 29.96 2.19 1.04 2. Allows for a unique solution for each pixel once a gamut class has been chosen.67 6.06 4. Removes the nonuniqueness (degeneracy) problem due to the use of a 3-to-3 inversion by selecting appropriate sub-gamuts.35 5.68 29. in-gamut color sweeps can be rendered without contours.54 29.31 Mean 1.82 4.49 4.21 95% 2. Hence.35 1.378 Control of Color Imaging Systems: Analysis and Design FIGURE 7. and CYK gamuts along the chromatic axes when viewed from the top of the gamut. the toner usage is usually smaller than that of other GCRs.18 1.96 6.

CMY. the node colors have to be carefully assigned to one of the CMK. MYK. it is expected that the appearance on the destination printer follows the proof. and CYK). Whereas. The overlap region with the CMK. and MYK gamuts are generally nonexistent except for numerical precision. The blended GCR response is shown in Figure 7.. 7.110] (i. or MYK gamuts and the overlapping CMY gamut should be avoided. CYK. yet another zero black tricolor (CMY) gamut class. Minimum black strategy would not work for images due to discontinuities or abrupt changes in separations.51).e.. MYK. as seen in the CMYK response for the colors on the neutral axis from C ¼ M ¼ Y ¼ K ¼ 0 to K ¼ 255 corner (C ¼ M ¼ Y ¼ 0 and K ¼ 255) (Figure 7. and CYK gamuts.50).49 Top view with CMK. For maximum black strategy.5 GCR RETRIEVAL FROM HISTORICAL PROFILES In this workﬂow. CYK.52. no CMY gamut) including neutrals (Figure 7. use black everywhere (i. CMK. for rendering spot colors.Three-Dimensional Control of Color Management Systems 379 CYK MYK CMK FIGURE 7. While using this GCR.9 (multidimensional ﬁltering and control-based iterations multiple times) of a four tricolor gamut class GCR [109. it can lead to severe formulation jumps and discontinuities unless some form of iterative blending is introduced. a minimum black strategy with CMY gamut class can be used without worrying about discontinuities. particularly in the prepress and=or in digital front end (DFEs). the documents are designed using various layout tools and their color appearance is ﬁne tuned by typically prooﬁng on a workgroup digital printer or the press itself. A smooth GCR can be obtained for a minimum black strategy by using the steps of Example 7. Otherwise. If it does not follow the proof. then adjustments are made in many . When the prints are made. this strategy has two components: (1) minimum black strategy and (2) maximum black strategy. Whenever there is an overlap with a CMY gamut class.e. It overlaps with other classes that include the K separation.5.

a maximum . The GCR methods ﬁne-tune the use of CMYK separations for improving the appearance.. places. Sometimes.380 Control of Color Imaging Systems: Analysis and Design Corners = 0 0 0 255 250 K C M Y 200 CMYK values 150 100 50 0 0 10 20 30 40 50 60 70 From 0 to gamut corner 80 90 100 FIGURE 7.50 Example of a CMYK response to colors on neutral axis for three gamut class GCR (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)).51 CMYK response to colors on neutral axis for four gamut class GCR with CMY gamut as the fourth gamut class (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)). including the color management proﬁle LUTs. black in ﬂesh tones and sky tones) need less black.g. Corners = 0 0 0 255 250 K C M Y 200 CMYK values 150 100 50 0 0 10 20 30 40 50 60 70 From 0 to gamut corner 80 90 100 FIGURE 7. One of the key adjustments is the selection of GCR methods. In particular. some of the key colors (e.

52 CMYK response to colors on neutral axis for four gamut class GCR after two iterations of preconditioning steps (‘0’ on the x-axis corresponds to white (C ¼ M ¼ Y ¼ K ¼ 0)). or a low K GCR.. etc. when the errors are near zero).g.. Direct use of a GCR-embedded proﬁle as an initial CMYK LUT may produce contours and other undesirable image artifacts since the proﬁle LUT may not represent the inverse of the printer in which the GCRs have to be retrieved. ICC proﬁles). the GCRs can be reproduced with reasonable precision as compared with the desired GCRs. record new CMYK values corresponding to each node colors.41) based on the Jacobian matrix extracted from the printer model at the CMYK values of interest. The preconditioning method includes the following steps: 1. K GCR is preferred over either a medium K GCR. Extract the CMYK values for all the node colors (L*a*b* values) from the GCR contained input proﬁle. and a future reprint may be expected. Below we show a method to reproduce images by retrieving GCRs from different sources. Using this approach. then the color in the future reprint has to be retrieved from the original electronic documents with as much original GCR design as possible while making sure that the embedded or the associated graphics are intact. Hence. 2.e. This step involves using an appropriate controller in the loop (as in Figure 7.Three-Dimensional Control of Color Management Systems Corners = 0 0 0 255 250 K C M Y 381 200 CMYK values 150 100 50 0 0 10 20 30 40 50 60 70 From 0 to gamut corner 80 90 100 FIGURE 7. These adjustments are stored as standard proﬁles (e. . which otherwise would have appeared. After convergence (i. Run the 4-to-3 control iterations for each node color using the appropriate printer model or printer using the GCR contained in the input proﬁle. in this section we show a preconditioning=processing method that is applied to a proﬁle. When a print job has already been produced. resulting in expected image quality appearance in documents without contours or other undesirable image artifacts.

382 Control of Color Imaging Systems: Analysis and Design 3. 7. store new CMYK values corresponding to each node color in the starting LUTs for use in the run-time algorithm. This will provide suppression of the scum dots. After convergence. Such an appearance would be unacceptable for high quality color rendition. the number of CMYK combinations that can produce the same color is high near neutrals but low near the boundary of the gamut. due to print engine drifts. 4.. Figure 7.53 CMYK separation without the use of a K-suppression algorithm for neutrals. There is only one unique CMYK solution on the surface of the gamut. during the updating process in the ﬁeld where the press would operate. Filter the CMYK values generated in step 2 using multidimensional ﬁltering algorithms [108]. . Most visible for neutrals.6 K-SUPPRESSION METHODS Now. The gain matrix shown in Figure 7. 250 Cyan Magenta Yellow Black CMYK response to neutral 8 bit input 200 150 Black can cause graininess. if required.g. this 4-to-3 control approach can inject black toner in regions where it is not needed (e. Rerun steps 2 and 3.41 (for each node color) is designed with LQR to minimize a selected quadratic objective function over the iteration length. scum dots in neutral ﬂesh tones or excessive black in the ﬂesh tones) even after designing good GCRs with no black in the initial CMYK LUT. that is.5. sky tones and other important tone scales can appear dirty=grainy or nonuniform. if required.53 illustrates a quantitative assessment of the problem where scum dots for neutral ﬂesh tones or excessive black in ﬂesh tones. in a computer program. This is particularly enhanced at the neutrals because of the high degree of degeneracy. 100 50 0 0 <----Black 50 100 150 R = G = B 8 bit input (a* = b* = 0) 200 250 White----> FIGURE 7.

e. by using the LQR design. can allow the K to be suppressed (or not). we are able to automatically provide low gain for black dots in the regions where black is not desirable. Varying the scale value.. The linear quadratic controller (Section 5. when the user ﬁnds excessive black in neutrals. r.Three-Dimensional Control of Color Management Systems 383 Construction of the gain matrix: We use the LQR for designing the gain matrix of the feedback controller that is used to compute the CMYK values at the proﬁle nodes for the updated printer model. Figure 7. The new gain matrix uses the weight proﬁle to emphasize the removal of black dots via the appropriate values in R matrix. Generally.3.2) minimizes a selected quadratic objective function for a node color over the iteration length.41 is used for obtaining the gain matrix with LQR design. N.. We state the ﬁnal equations below. they can change the values of the parameter r. . The gain matrix equation is obtained by using the procedure described in Chapter 5. q1 ¼ q2 ¼ q3 ¼ 1 Â 10À3).54 illustrates the neutral response without the scum dots. Since our problem is focused on suppressing black.g. Gain matrix: Â ÃÀ1 K(k) ¼ RÀ1 BT P(k þ 1) I þ BRÀ1 BT P(k þ 1) A Recursive equation to compute P(k): À Á P(k) ¼ AT P(k þ 1)A À AT P(k þ 1)BRÀ1 I þ BT P(k þ 1)BRÀ1 Â BT P(k þ 1)A þ Q Boundary condition: P(N) ¼ 0 (7:94c) (7:94b) (7:94a) It turns out that the state space model for each node color has an A matrix which is equal to an identity matrix (i. which is shown below J¼1 2 NÀ1 P k¼0 xT (k)Qx(k) þ uT (k)Ru(k) (7:92) where x(k) is the state vector containing L*a*b* values and u(k) is the actuator vector for four-color system. Thus. The state space formulation of Figure 7. The R matrix contains a as the weight which is used to suppress black. Note that k refers to iteration number in the above equation. A ¼ diag[1 1 1]). we set the values in the Q and R matrices as follows: Q ¼ diag½q1 R ¼ diag½r1 q2 r2 q3 r3 a (7:93a) (7:93b) The Q matrix is 3 Â 3 with very small ﬁxed values for the elements (e.

right: side view).55). bright secondary colors (such as FIGURE 7. Colors present in the source gamut but not in the output gamut are said to be ‘‘out-of-gamut’’ and must be accounted for before the digital image can be printed.54 CMYK separation with the use of a K-suppression algorithm. That is. . For example. The typical computer monitor has a wider gamut than the typical color printer (see Figure 7. areas in the output document where these colors are found cannot be simply left blank in the ﬁnal printed document. solid: printer gamut.55 (See color insert following page 428. 7.) Comparison between display gamut and a typical color printer gamut (wire: sRGB display gamut.384 250 CMYK response to neutral 8 bit input Control of Color Imaging Systems: Analysis and Design 200 Cyan Magenta Yellow Black 150 100 No black dots 50 0 0 <----Black 50 100 150 200 250 R = G = B 8 bit input (a* = b* = 0) White----> FIGURE 7.6 GAMUT-MAPPING METHODS The printer gamut is generally limited when compared to the gamut of the source digital image due to the physical limitations of the printer’s colorants. left: top view.

both in-gamut and the out-of-gamut colors are altered in order to map the entire range of image colors to the printer gamut. the gamut of the device in which the image will be reproduced).e. these colors cannot be reproduced on the printer and would simply be clipped. the bright cyan of an inkjet printer may not be easily presented on an average computer monitor.1 GAMUT MAPPING WITH RAY-BASED CONTROL MODEL All gamut-mapping methods need to determine accurately whether the node color is inside or outside the destination gamut. It is a key feature used in every color reproduction device. In other words. hue. There are at least two gamuts involved in mapping colors: (a) source=image gamut (i. device designers generally compromise regarding the gamut-mapping functions they wish to employ in their respective color management systems. contrast. and (3) printer gamut. Techniques used for dealing with out-of-gamut colors include gamut clipping and gamut compression. there will be three gamuts involved: (1) display=monitor gamut.e.. Some gamut-mapping algorithms offer feature enhancements in one region of the gamut and others are more favorable elsewhere in the gamut. lightness. gamut of the image whose colors need mapping) and (b) destination gamut (i. Conversely. 7.Three-Dimensional Control of Color Management Systems 385 lime green) that appear on the display monitor cannot be printed using subtractive process colors. A color that is wrongly considered as outside . (2) source= image gamut. Particularly when displays are involved. In gamut clipping. As such. For efﬁcient computational processing of images in practical systems. There are several techniques used for gamut mapping [111–133]. as is normally the case when soft prooﬁng images. while in-gamut colors are left unaltered. wherein a ray is drawn from a desired outof-gamut color to a point on a neutral axis. In addition to this problem. chroma. The out-of-gamut colors are converted to printable colors through a transformation called gamut mapping or gamut compression. and the like across all imaging devices. The location or point where the ray penetrates the gamut surface is the gamut-mapped color. A common form of clipping involves a ray-based approach.. In gamut compression. Otherwise. gamut mapping is a means of tone scale modiﬁcation that attempts to preserve the original appearance of a color image captured with a wider gamut device on a smaller gamut device. This is true for pictorial images as well as business graphics and computer-generated images. a gamut-mapping algorithm that produces pleasing results for one image may not work well for another. There is no unique gamut-mapping method that satisﬁes all requirements for image reproduction. The blue color printed on paper with CMYK toners=inks may not be as saturated as the one seen on a monitor. For instance. all out-of-gamut colors are mapped to a color on the gamut ‘‘surface’’ in some way that minimizes the degradation of the resultant output. Such a strategy is implemented to preserve hue through the gamut-mapping operation. such as pleasing color. the gamut-mapping operation is incorporated into the nodes of a 3-D LUT.6. producing a mostly saturated blue color as displayed on a monitor as a printed output on paper (using a typical CMYK printer) will likely fail. The color management system can utilize various gamut-mapping methods to achieve the desired results and give experienced users control of the gamut-mapped outcome.

g. as in Figure 7. Such mistakes can easily be avoided by not mapping the wrongly assigned in-gamut color..95 with start and end points given by x0 and xc. A GCR LUT can b 127 xc xb –128 x0 a 127 –128 Gamut in 2-D chroma plane FIGURE 7. expressed in L*a*b* space by Equation 7.56 shows schematically a chroma plane and a line starting from point x0 to the node color xc. In such a case. Consider a ray (modeled by a line). Figure 7. When i ¼ 0. are carried out using x value as the target for i ¼ 0 to N.. the line will end on the node color.g. a centroid of the gamut).386 Control of Color Imaging Systems: Analysis and Design the gamut may be handled incorrectly by the gamut-mapping algorithm. b 3 2 3 L0 L À L0 14 c x0 ¼ 4 a0 5. then the consequence of this decision might lead to mapping the colors to a wrong region of the destination gamut. We present below a control-based method to decide whether node colors are inside or outside a printable gamut. x0 is inside the gamut (e.95. where N is equal to the number of points along the line. contours) and color errors when images are rendered through those multidimensional LUTs. m ¼ ac À a0 5 N b0 bc À b0 2 (7:96) (7:95) xc is the node color whose in–out determination has to be made. In Equation 7. . there will be a total of N þ 1 colors whose values are given by x along the ray between x0 and xc (including extreme points). the reproduction of the mapped color can be ‘‘way off’’ from the original. for spot colors. Now. which can lead to unacceptable artifacts (e. When i ¼ N.41. Also. respectively: x ¼ x0 þ mi where 2 3 L x ¼ 4 a 5. where x0 and xc have coordinates {L0 a0 b0} and {Lc ac bc}. 4-to-3 control iterations. i is the index that is incremented from 0 to a convenient integer until the node color xc. For i ¼ 0 to N. the line will start at x ¼ x0.56 A line between start and end points. if the outof-gamut colors are wrongly assigned as ‘‘in-gamut’’.

02 0.25 0.03 0.34 81.09 0.04 0.15 1. In Table 7.10 2.18 76.80 58.97 81.64 94.33 7.73 1.81 Y 70.15 87.90 0.13 0.76 42.01 0.40 5.80 5.46 6.01 0.94 2.03 0.50 11.80 2.46 7.02 0.02 0. ac ¼ 50.26 108.53 82.02 1.64 0.15 0.17 0.60 50.06 77.07 0.00 8.16 0.56 72.72 72.02 1.74 72.50 8.60 3.00 3.36 0.07 0.19 0.53 75.13 0.02 0.16 0.40 55.00 50.02 0.00 8.06 0.09 0. If the DEab numbers are close to zero (or close to a predeﬁned threshold value.53 2.00 11.31 3.29 103.66 1.05 0.17 73.30 69.18 3.20 1.01 0.01 0.03 0.08 111.01 0.94 0.93 2.00 0.15 1.51 1.29 0.34 81.44 0.35 77.60 6.49 0.18 0.80 0.13 7.50 0.06 0.02 0.01 0.97 97.00 0.00 0.00 0.86 2.00 0.00 10.00 5.80 2.71 94.03 0.43 3. Color numbers 0–76 are considered as being inside the gamut.02 0.02 0.30 3.04 0.45 75.88 100.39 75.05 0.00 53.26 1.70 0.00 9.59 103.99 7.00 0.33 0.99 2.02 0.17 0.00 16.22 8.79 0.00 70.04 83.18 0.11 Example Illustrating the Convergence of a Ray between x0 ¼ [L0 ¼ 50.67 1.63 92.30 0.03 0.38 68.32 89.69 2.49 3.04 0.79 72.18 93.90 54.30 56.57 85.01 0.01 0.01 0.01 0.70 89.19 1.00 2. then those colors are considered inside the printer gamut.43 115.91 45.50 13.00 0.13 0.30 59.63 81. bc ¼ 30] for Colors between 0 and 50 Iteration# Color# 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 L * Best CMYK C 87..08 0.12 0.51 2.89 118.29 92.22 6.36 0.19 0.88 66.35 0. each with 10 iterations.90 110.04 0.49 1.14 79.00 0. color numbers 77–100 are out-of-gamut.49 7.77 2.23 97.01 0.89 0.62 79. e.00 13.80 8.01 0.43 80.38 60.84 114.70 3.04 0. then we can say that those colors are out-of-gamut.62 78.03 0.04 0.54 1.80 0.15 107.26 70.30 0.09 0.22 87.46 1.20 51.00 59.48 M 70.00 0.10 8.11 and 7.35 0.01 119.90 7.73 112.01 0.73 0.50 14.80 55.01 0.68 82.50 5.60 6.94 0.39 43.67 8.09 74.01 0.77 8.11 0.65 1.02 0.07 0. a0 ¼ 0.68 7.90 4.38 0.70 9.06 0.71 85.01 0.74 72.60 53.22 0.60 0.00 7.70 74.18 75.39 7.10 0.04 0.00 0.00 9.75 0.00 0.00 0.18 0.39 3.05 6.01 0.04 0.37 3.01 0.40 0.01 0.76 121.06 8.15 96.01 0.36 0.45 0.94 0.88 0.94 86.98 73.12 show an example of * * the DEab convergence numbers for 101 colors.93 83.00 50.00 0.01 0.73 0.07 0.12 1.00 0.60 0.40 7.60 59.66 1.05 0.67 71. 0.08 0.00 0.37 82.49 0.00 0.36 1.50 9.07 0.12.10 52.18 7.10 0.09 0.10 55.01 0.01 0.03 78.91 99.09 0.01 0.59 83.04 0.64 0.12 70.83 116.56 82. The Euclidean norm * (DEab) is computed for each iteration.89 91.75 120.00 2.02 0.20 0.09 0.30 0.14 55.01 0.90 57.00 0.00 14.90 122.29 0.00 0.90 1.10 5.00 0.86 72.04 7.05 0.63 7.82 76.50 4.82 63.64 1.02 0.01 0.55 1.30 6.74 2.57 89.00 10.00 6.09 0.21 84.20 7.00 4.64 67.01 0.00 0.59 1.08 0.16 0.52 75.00 0.07 0.35 0.00 6.15 96.75 0.40 8.04 0.73 92.85 81.98 2.09 0.00 0.29 3.00 12.00 0.48 1.50 2.97 3.20 54.42 0.00 0.35 0.48 3.00 0.86 86.07 0.00 3.04 73.02 0.01 0.02 0.51 88.75 82.00 0.02 0.32 0.08 7.01 0.17 0.51 100.74 53.13 100.75 2.20 57.72 2. b0 ¼ 0] and xc ¼ [Lc ¼ 80.41 97.18 0.02 0.40 2.79 0.18 0.20 7.36 78.81 77.99 1.05 0.94 106.40 52. TABLE 7.12 0.21 80.04 0.01 0.00 4.40 0.60 84.50 3.30 9.01 0.59 2.04 50.39 0.01 0.17 7.01 0.04 0.50 16.g.25 80.07 0.04 0.79 76.00 7.63 2.50 57.49 3.50 7.00 6.03 0.00 0.74 2.01 0.16 0.17 0.05 0.37 0.01 0.07 0.09 71.50 51.05 0.80 0.03 0.00 6.06 70.66 3.11 0.70 53.02 99.38 0.16 0.00 3.00 0.Three-Dimensional Control of Color Management Systems 387 be used to extract the nominal CMYK values for a given x value.02 1.01 0.33 89.05 0.03 0.10 93.00 0.79 2.20 7.26 58.09 2.30 101.43 87.00 56.09 74.80 52.00 0.41 0.52 K 98.65 3.00 0.03 0.75 84.00 0.40 86.37 0.50 7.19 0.26 0.03 0.61 7.50 1.29 67.29 1.43 0.82 7.01 0.50 6.93 96.50 54.00 1.61 0.16 1.30 50.00 0. Tables 7.57 0.40 58.50 15.12 91.39 0.66 8.94 8.05 0.27 90.50 12.02 0.03 0.00 0.90 7.09 0.88 94.60 56.85 8.94 0.93 57.60 9.05 98.08 0.00 0.70 6.70 59.01 0.70 56.20 4.45 0.01 0.10 58.60 3. * If the DEab numbers are larger than the threshold value.01 0.27 98.40 83.13 0.00 5.00 8.05) after steady state is reached.70 2.92 62.17 85.02 0.01 0.00 0.31 0.21 6.90 (continued ) .16 0.61 95.50 4.71 0.97 8.33 0.00 15.50 1.45 a * b * 1.20 46.19 0.50 10.30 95.00 3.90 51.15 0.21 7.55 0.03 0.65 49.02 0.79 52.26 3.00 0.59 78.01 0.95 1.00 9.94 93.23 0.30 53.

01 0.90 13.20 0.15 131. M ¼ 169.93 0.60 12.78 4. M ¼ 150. bc ¼ 30] for Colors between 0 and 50 Iteration# Color# 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 L * Best CMYK C 66.63 109. M ¼ 148.61 128.01 0.15 0.40 140.85 5.25 141.49 108.50 21.13 0.94.44 4.10 14.34 2.388 Control of Color Imaging Systems: Analysis and Design TABLE 7.99.00 5. in Table 7.22 107. these values represent the best CMYK values for minimum DEab (for both in & out-of-gamut colors)..93 125.12 1. In * other words.45 131.52 5. (2) can be gamut mapped with minimum DEab with a gamut-mapped CMYK value of C ¼ 0.00 20. when the node color is projected along the ray in the ray-based gamut mapping.88 22.00 2.40 11.75 60.05 0.05 0.00 17.42 57.50 4. it is also easy to ﬁnd the mapped CMYK values for the node color.00 8.02 0.01 0.49 5.95 0.01 0.06 0.02 0.52 0.02 0.19 0.01 0.00 0.93.01 0.01 0.01 0.01 0.02 0.18 136.30 12.27 0.30 62.00 6.28 2.03 0.50 20.00 9.40 61.89 0.00 0. ac ¼ 50.49 14.40 14.96 4.03 0.90 63.28. The CMYK values of the nearest out-of-gamut node give the gamut-mapped CMYK values for the node color xc when the mapping is done along the ray.00 0.73 59.04 137.16 2.28 2.70 65.46 127.11 0.50 24. xc.43 0.81 105.38 0.44 139.00 The CMYK values achieved during convergence are also shown in last four columns corresponding to each row.04 107.00 24.10 64.23 5.84 65.26 105.00 0.52 64.28 37.00 23.00 0.89 0.00 3.01 0. Y ¼ 103. a0 ¼ 0.20 13.50 10.83 63.25 58. the node color.80 14.00 0.06 0.26 33.00 60.01 0.09 107.90 19.89 2. Y ¼ 110.13 0.00 10.03 2.10 101.06 0.03 0.50 60.16 2.70 12.03 0. xc can be mapped to color# 72 with a gamut-mapped CMYK value of C ¼ 7.82 0.57 135.72 129.94 1.00 4.06 1.00 17.15 0.51).15 4.01 0.76 133. (3) can be gamut mapped along the ray on the gamut surface with a gamut-mapped CMYK value of C ¼ 0. They are the best CMYK values for the chosen * GCR used to determine the nominal CMYK that give the minimum DEab error.19 1.02 0.26 0.76.01 0.00 17.07 0.00 0.81 M 123.85 39.46 0.02 0.22 2.50 25.14 5.16 0.22 0.00 0.87 60.50 19.70 62.01 0.02 0.77 108.11 0.63 103.79 DEab.50 24.08 35.50 13.00 0.99 1.92 5.50 63.00 7.14 105.24 62.57 28. and (4) can be mapped inward along the ray as in compression methods (e.62 56.49 0.90 0.01 0.39 29. K ¼ 2.06 0.00 18.02 0.50 22.23 108.17 2.61.37 0. This is obtained when the ray penetrates the surface.70 65.50 18. Y ¼ 103.02 0.50 23.17 2.02 0.98 0.10 11.80 61.02 0.53 0.00 0.06 0.80 61.20 10.24 0.11 (continued) Example Illustrating the Convergence of a Ray between x0 ¼ [L0 ¼ 50.03 0.07 0.39 2.98.00 0.43 K 40.13 0.38 0.06 0.92 0.00 0.07 0.12 (see marked boxes).05 0.26 2.10 54. xc (1) is out-of-gamut * * by 15.13 0.04 0.02 0.g.46 12.05 15.06 0.47 0. K ¼ 0.57 102.00 0. From this table. K ¼ 0. As is the case in gamut clipping method—one of the ray-based approaches the gamut-mapped color will be on the surface when this approach is used.31 2.12.01 0.46 55.33 0.01 0.00 10.35 2.38 0.17 0.00 22.38 0.84 132.12 0.26 0.14 0.70 a * b * 1.37 4. A ray-based control model can also be used to perform an accurate linear mapping for the compression algorithms.25 0.03 0.08 1.36 0. In gamut compression method—another ray-based approach—the gamut-mapped colors are moved inward along the ray to the nearest node with a scaling associated with the local region of the gamut.00 0.74 5.95 0.52 0.02 126.10 61.06 0.43 103. .69 Y 101. Thus.00 19.74 5. say for example.13 2.00 5.40 0.07 0.01 0.70 106.40 64.25 26.98 0. It is equal to 77th color in Table 7.42 0.20 63.26 0.10 20.06 0.06 0.15 0.05 0.00 21.70 15.09 0.09 0.25 4.52 32.05 103.02 0.00 0.28 138.38 64.13 0.02 0.02 0.60 62.00 62.01 0.03 0. b0 ¼ 0] and xc ¼ [Lc ¼ 80.04 0.80 11.15 0.08 0.20 60.55 0.76 4.00 12.80 64.01 0.02 2.85 65.

07 0.41 0.63 147.77 0.00 27.00 7.17 0.18 0.00 0.02 0.36 0.17 0.26 33.43 0.01 0.53 1.10 6.31 145.04 0.00 71.33 2.60 65.16 107.08 0.39 107.00 C M Best CMYK Y 109.00 53.01 17.90 16.94 44.63 146.22 0.29 3.04 0.00 0.79 1.23 2.00 0.30 18.81 109.23 0.93 0.00 5.47 0.35 9.62 5.00 30.44 8.02 0.31 8.07 2.10 70.85 149.05 0.60 2.12 39.58 149.00 29.25 21.73 1.79 0.00 0.82 3.33 0.01 0.00 8.04 0.62 3.05 108.25 147.50 26.40 70.60 18.64 14.00 0.78 0.00 0.00 0.13 10.70 18.03 0.92 0.00 0.67 148.85 0.55 3.00 0.83 0.70 68.83 0.01 0.40 0.00 0.70 71.00 21.50 31.00 0.81 1.02 2.81 0.05 0.00 0.20 16.83 148.12 Example Illustrating the Convergence of a Ray between x0 ¼ [L0 ¼ 50.09 0.38 0.08 0. b0 ¼ 0] and xc ¼ [Lc ¼ 80.92 4.80 17.25 3.02 0.53 109.18 0.07 0.84 0.02 0.01 0.90 66.75 1.17 0.39 0.47 31.03 0.00 0.35 25.01 0.82 0.15 42.43 2.00 0.20 0.80 70.72 4.28 3.57 1.50 16.39 0.61 104.32 4.60 68.40 0.04 0.08 0.50 109.20 19.89 148.00 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 65.04 0.50 33.18 0.26 0.04 0.80 67.44 3.03 0.81 4.36 0.01 0.31 0.90 109.14 0.00 0.83 0.97 5.41 0.03 3.40 67.94 1.82 4.80 K 11.00 1.88 0.24 4.02 0.73 109.08 0.02 0.01 0.15 0.62 107.24 148.11 148.00 0.40 17.30 4.03 0.04 0.98 0.82 144.00 10.01 0.01 0.71 50.27 2.50 30.12 3.06 0. bc ¼ 30] for Colors between 51 and 100 Iteration# 3.46 23.01 0.54 4.15 0.09 0.34 0.72 0.13 0.09 0.91 37.34 0.70 145.03 0.97 1.01 0.03 109.12 143.09 0.00 68.57 146.01 0.01 0.76 148. a0 ¼ 0.50 34.87 4.75 11.10 0.02 0.15 0.02 0.06 0.00 1.86 108.60 15.50 35.10 67.90 69.50 29.00 0.70 109.02 0.15 2.80 3.09 0.12 4.40 20.76 3.11 0.02 0.00 2.01 0.66 109.00 18.50 28.TABLE 7.34 105.01 0.30 25.01 0.50 69.14 4.05 1.02 0.90 19.00 0.19 0.10 3.67 1.33 4.10 20.50 32.20 69.02 0.38 0.01 0.96 147.00 0.01 0.00 26.95 3.00 0.04 0.01 0.79 0.00 0.00 0.58 148.58 148.00 9.50 15.47 4.48 109.38 44.02 0.80 142.14 106.00 0.02 0.40 0.50 27.02 0.09 0.97 47.30 65.00 33.97 (continued) Three-Dimensional Control of Color Management Systems Color# L* a* b* 1.00 32.09 0.50 19.13 0.02 0.15 0.00 6.05 0.78 1.67 0.57 28.03 1.48 110.01 0.10 17.06 0.74 1.04 0.00 31.61 104.92 1.70 21.00 0.89 10.66 51.19 0.01 0.20 3.95 1.00 35. ac ¼ 50.08 0.62 143.61 11.91 1.00 28.79 0.76 2.77 389 .34 0.04 0.30 15.91 35.00 0.63 4.17 0.80 20.04 0.00 34.01 0.85 0.57 0.01 0.78 2.16 48.04 0.30 68.98 108.36 0.02 0.19 0.01 0.20 66.50 66.01 0.00 0.

50 72.60 2.41 154.00 39.10 2.02 0.28 103.88 4.80 73.00 0.07 155.30 74.30 24.48 4.20 22.72 3.83 2.50 2.28 3.61 2.00 0.60 71.00 0.01 0.69 3.01 0.00 Control of Color Imaging Systems: Analysis and Design Color# L* a* b* 1.50 2.46 103.80 23.34 0.32 3.54 2.09 0.65 Y 103.41 0.22 0.94 5.59 3.00 0.10 0.34 1.46 5.71 3.50 42.08 0.60 24.51 2.50 2.70 4.70 3.79 3.00 41.01 0.49 0.03 1.00 0.00 21.10 2.00 38.95 4.06 1.49 0.00 Best CMYK M 148.00 0.93 103.32 0.31 1.01 0.00 0.59 102.09 0.90 4.71 1.10 73.62 0.18 149.51 0.89 4.00 6.20 3.60 21.87 4.44 1.50 0.20 1.00 40.69 3.32 3.00 74.60 74.52 0.51 2.87 4.00 0.38 1. b0 ¼ 0] and xc ¼ [Lc ¼ 80.00 24.390 TABLE 7.40 73.70 24.00 7.00 0.02 0.33 0.11 2.70 3.40 23.00 0.71 2.80 152.08 5.01 0.03 0.95 1.00 72 73 74 75 76 77 78 79 80 81 82 83 84 71.58 0.00 8.03 0.30 3.00 0.99 149.50 22.20 0.23 0.98 151.52 2.74 3.96 103.31 3.91 1.01 0.17 104.70 .91 1.00 0.03 4.00 2.53 0.00 0.02 3.28 2.50 37.92 1.61 103.00 0.33 1.18 150.20 0.70 74.41 0.47 10.00 37.20 36.73 1.00 0.91 1.50 41.01 0.25 155.38 4.47 C 7.22 3.91 1.01 0.24 0.98 3.10 0.80 0.08 0.00 0.00 0.90 156.51 2.29 3.50 39.10 150.91 1.89 105.62 153.27 4.90 75.19 0.43 4.10 2.00 0.05 0.90 22.52 0.32 0.25 105.44 0.16 4.14 2.58 103. a0 ¼ 0.56 0.50 40.12 (continued) Example Illustrating the Convergence of a Ray between x0 ¼ [L0 ¼ 50.00 0.96 4.20 72.49 K 2.31 3.10 2.00 0.00 0.23 1.04 0.10 23.50 38.55 150.28 3.90 104.00 0.61 1.00 0.66 4.11 2.05 0. bc ¼ 30] for Colors between 51 and 100 Iteration # 3.53 104.00 0. ac ¼ 50.32 0.00 0.12 2.00 9.00 0.00 5.00 36.92 4.04 0.90 72.90 25.87 2.36 3.

00 0.15 8.00 0.70 30.52 158.00 0.86 7.57 162.85 10.17 6.09 12.98 13.08 13.00 0.37 12.95 12.91 10.00 77.91 165.50 50.26 6.21 12.24 106.80 29.35 15.45 109.76 105.00 6.46 168.50 75.06 8.00 0.50 25.00 0.87 7.96 7.00 0.22 12.23 9.66 110.25 6.79 13.80 13.00 46.81 9.37 14.85 7.20 107.50 46.00 0.61 0.00 0.45 13.47 167.00 157.03 11.22 5.10 169.Three-Dimensional Control of Color Management Systems 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 75.86 8.51 11.10 8.47 163.12 11.55 10.05 8.48 11.11 164.63 8.85 12.00 0.86 13.12 8.83 13.25 12.78 13.70 12.98 5.30 6.90 28.23 12.27 5.65 6.28 12.11 5.44 110.64 7.00 44.71 9.75 9.64 9.02 5.80 26.80 76.50 45.34 159.02 11.04 9.03 10.35 11.40 26.68 109.82 10.60 27.59 7.17 11.82 10.00 0.00 0.83 8.72 13.10 5.00 49.81 7.26 6.50 48.00 0.45 108.15 5.60 77.50 78.20 169.42 6.09 8.06 5.50 28.40 76.68 9.50 49.49 8.20 78.55 11.00 0.25 10.00 0.43 11.35 6.42 11.45 8.66 9.00 0.00 0.10 26.50 47.27 11.92 7.05 11.00 5.91 13.26 9.42 6.79 108.76 6.96 10.37 107.02 14.84 9.76 12.11 110.70 14.00 391 .66 12.78 107.72 6.00 0.00 0.40 29.00 0.00 0.69 6.82 6.40 14.32 12.00 43.63 12.97 0.90 7.00 25.02 108.00 0.90 78.03 6.66 6.59 14.46 11.01 11.00 0.55 15.10 29.07 5.08 11.55 12.06 5.43 8.00 0.05 8.11 14.08 5.79 13.89 106.37 13.07 8.17 5.61 11.07 5.08 10.10 79.00 27.59 12.38 13.00 0.85 7.24 9.61 12.12 109.51 11.52 106.20 159.47 8.63 9.00 0.83 11.00 42.00 0.56 8.84 161.30 27.84 166.12 5.00 48.65 6.43 14.44 8.40 79.79 5.27 10.54 5.88 7.00 47.02 14.67 6.00 0.32 6.44 11.89 160.70 77.52 8.80 79.65 9.50 43.52 14.20 28.70 80.28 6.50 44.28 9.47 14.32 9.84 10.30 9.70 27.10 7.10 76.30 77.25 9.00 0.20 164.00 0.60 12.37 10.33 9.00 45.88 10.

1. chroma. a* ¼ 6. 7.57 Rays pointing to the focal point in centroid gamut mapping. which are overlaid on top of the actual * printer gamut. L*. C. Next we show how to apply the ray-based control model to perform few wellknown gamut-mapping strategies. either for minimum DEab mapping (also called nearest point clipping) or for centroid clipping to the gamut surface. b* ¼ 13.) using ray-based control model to provide variations to the components such as lightness.2 CENTROID CLIPPING In centroid gamut mapping.6. the DEab convergence value during iteration deviates to greater than zero at the gamut intersection as shown by arrows.6.11). Figure 7.e. it is straightforward to use with ray-based control model to map colors accurately. can be easily obtained using the procedure outlined in Section 7. .392 Control of Color Imaging Systems: Analysis and Design Thus. H. Additional constraints can be included during the iterations while performing error minimization (i. but also performs accurate gamut mapping along the ray axis. we use the method that maps to the gamut surface along the ray toward the gamut centroid (x0)—the focal point.57 shows rays pointing from the sample node colors to the centroid (L* ¼ 60.62. and hue. Though this type of gamut mapping reduces chroma. Each ray was * divided into 101 points. Figure 7. In our illustration. Mapped * CMYK values.. FIGURE 7. Clearly. the node colors with a constant hue are mapped toward a focal point inside the gamut [131].13.58 shows the DEab convergence values at the end of 10 iterations for six out-of-gamut node colors shown in Table 7. DE2000 etc. this control-based iteration of the ray model not only gives an accurate determination of whether or not the colors are located inside or outside the gamut surface.59.

12 68.31 7.Three-Dimensional Control of Color Management Systems 140 Mapped * ΔEab values with respect to node color 393 120 100 80 60 40 20 0 0 10 20 30 40 50 60 70 Color number along the ray 80 90 100 * FIGURE 7.00 À0.19 À1. x0. TABLE 7.50 À122.88 b* 0.58 Out of Gamut Node Colors 1 2 3 4 5 6 L* 0. one of which is the soft gamut mapping illustrated below. xc.58 Converged DEab values shown for six node colors.6.12 À0.13 Out-of-Gamut Color Considered in Figure 7.00 À54. This approach is shown graphically in Figure 7.72 À1.55 À150. xc could be selected as the color on the boundary of the source color space. Let xb be the nearest out-of-gamut color along the ray between the in-gamut color.22 88.3 SOFT GAMUT MAPPING WITH RAY-BASED CONTROL MODEL A ray-based control model can be used to implement accurate gamut compression algorithms. All the out-of-gamut colors between x0 and xc are mapped to the region between x0 and xb nonlinearly inside the gamut surface using a smooth curve.75 a* 0. For this discussion.00 7.36 À0.59 46.08 À175.59 with a dashed curve. . and the out-of-gamut node color.12 À91.68 23.

the interval {x0 to xc} is mapped linearly into {x0 to xb}. All colors greater than xb are mapped to xb. the in-gamut colors are compressed less. all in-gamut colors are passed through the mapping without any compression. h h xb À x0 i xc À x0 i i þ b x0 þ i for 0 x ¼ (1 À b) x0 þ N N h xb À x0 i x ¼ (1 À b) x0 þ i þ bxb for ib i N N i ib (7:97) (7:98) where 0 b 1 is the compression parameter. all colors including the in-gamut colors are moved inward except for the node color xc which is clipped to the boundary.60 shows L*.000. A conventional gamut-mapping approach is described below that preserves the lightness and hue angle. For b ¼ 1.5.2 with a set equal to 100.97 and 7. With b ¼ 0. Figure 7.3.6. b* curves with respect to color indices i for three different values of b. Both b and a contribute to the overall smoothness of the soft gamut-mapping curves and can be customized differently to each region of the gamut.59 Soft gamut mapping. . As b increases to a value greater than 0. A soft gamut-mapping curve can be constructed using a piecewise linear compression function shown below followed by a 1-D smoothing algorithm described in Section 6. we used the 1-D ﬁltering algorithm described in Section 6. A raybased control model is run for b ¼ 0 and i ¼ 0 (representing color x0) to i ¼ N (representing color xc which is now mapped to xb) to obtain the CMYK values corresponding to the colors along the ray. If the out-of-gamut node color happens to fall anywhere between xb and xc then it is mapped according to the rules described below.394 Control of Color Imaging Systems: Analysis and Design xc Mapped colors in L*a*b* xb Soft gamut mapping curve x0 xb Colors along the ray xc FIGURE 7. In Figure 7. When b ¼ 0.98.60.1. a*. Smoother compression can be achieved using the 1-D ﬁltering of the function shown in Equations 7. i is the index incremented between 0 to a convenient integer value until xc.

4 GAMUT MAPPING FOR CONSTANT LIGHTNESS AND HUE An algorithm that preserves the lightness and hue angle of the node Lab color coordinates is shown below. i β = 0.5. This is achieved by the following two steps: 1. First we need to determine whether a given input color is inside or outside the printer gamut. as illustrated in Figure 7. then. approximately equal to Lab ¼ [50 0 0]. 1 shown with and without smoothing. and b* (bottom center) for b ¼ 0.60 Soft gamut-mapping curves for L* (top left). the ‘‘*’’s are removed from the color coordinates. 7. If the original color in the CIELab color space is denoted by [L a b] and its corresponding mapped color by [L0 a0 b0 ]. a. In these new color coordinates. this can be shown on a chroma plane at constant L* by drawing a ray from the desired color to the centroid of the gamut. each node color is speciﬁed by its spherical coordinates r. L*a*b*. Transform the color gamut into spherical coordinates where the center of the sphere is at the centroid of the gamut.Three-Dimensional Control of Color Management Systems 80 75 70 65 60 55 50 45 0 10 20 30 40 50 60 70 80 90 100 Color index.5 β = 0 a* 20 10 0 –10 0 L* β=1 50 40 30 395 10 20 30 40 50 60 70 80 90 100 Color index.61. i 30 25 20 15 b* 10 5 0 –5 0 10 20 30 40 50 60 70 80 90 100 Color index. i FIGURE 7.6. . a* (top right). to preserve the lightness and hue relationship. 0. and u given by the equations. the following constraints must be satisﬁed: L ¼L 0 and tan À1 0 b b ¼ tanÀ1 0 a a (7:99) Graphically. For convenience.

2. a is the hue angle with the dynamic range from 08 to 3608 and u is the angle in the constant a plane from 08 to 1808. Now the desired color point is inside the gamut if r0 < rave. all node colors are clustered in different regions of interest within the color space and each cluster is associated with at least one .99. Find the intersection of the gamut with a straight line connecting the input color point having coordinates [r0 a0 u0] to the centroid of the gamut. More than 90 gamut-mapping algorithms [111] can be used for mapping colors to a suitable region of the printer gamut. lightness.396 Control of Color Imaging Systems: Analysis and Design b 127 Lab L΄a΄b΄ α –128 127 a –128 Constant L FIGURE 7. r is the distance from the gamut centroid to the given color point. The average of these boundary points is denoted by [rave aave uave]. we map it to the nearest point on the gamut boundary that best satisﬁes the conditions shown in Equation 7. In this method. As such. chroma. This is achieved by searching the gamut boundary points with coordinates that fall into the range a0 Æ Da and u0 Æ Du. hue. Others do the same in different regions of the gamut. device designers generally compromise in the gamutmapping functions they wish to employ in their respective color management systems.5 MERIT-BASED GAMUT MAPPING There is no unique gamut-mapping method that can satisfy requirements like pleasing color. 7. Some offer feature enhancements in one region of the gamut. qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ (L À 50)2 þ a2 þ b2 b L À 50 and u ¼ tanÀ1 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ a ¼ tanÀ1 a a2 þ b2 r¼ (7:100) In these equations. etc.61 Gamut mapping for constant lightness and hue. Once we determine that the color is outside the printer gamut. contrast. otherwise it is outside the printer gamut.6. Merit-based feedback system used for gamut mapping offers a way to automatically select the gamut-mapping algorithms from the library of algorithms to optimize a merit function [134].

The DE2000 function calculates distances between the out-of-gamut L*a*b* values and the mapped L*a*b* values for each node color of the current cluster. a transformation block ‘‘T’’ maps the out-of-gamut L*a*b* values to a 1-D function for each color point.Three-Dimensional Control of Color Management Systems Gamut mapping algorithm (e. In this case. the merit function determines the mean-squared error of the values calculated using the DE2000 function for each node color belonging to a particular cluster that needs to be mapped. The optimized gamut-mapping functions thus obtained for adjacent clusters are then blended together using multidimensional smoothing algorithms or other methods described in Ref. and the one with the best merit function output can be chosen. b and a (or x0 in the centroid clipping algorithm). merit function value. A mean-squared error function can be used as the merit function. At least one merit function is associated with each cluster. The output of the merit function is a single value which represents the numerical merit of the mapping parameters. An optimization block manipulates the gamut-mapping parameters during successive iterations to generate the best. Gamut-mapping parameters are iteratively varied (or tuned) to optimize their values. candidate gamut-mapping function. In Figure 7. which is to say that a DE2000 mapping is being emulated. This could be.3. The entire process is done automatically in a computer program.62 Merit-based gamut mapping with closed-loop optimization. When the optimum merit function value is reached (or exceeded) the iteration loop ends for the cluster being processed. [134] to generate a composite smooth function that collectively exploits the local advantages of each cluster. A brute-force approach in many instances does not take much time to generate the ﬁnal results given present processor speeds. The merit function is selected based on what is required by the device designer during optimization. soft gamut mapping) 397 Mapped L*a*b* T 3 Merit function L*. b* Optimization algorithm FIGURE 7. The optimization algorithm could be a brute-force approach. A number of different candidate algorithms are iterated for each deﬁned cluster within the color space. with two gamut-mapping parameters.62) to iteratively tune various mapping parameters associated with the gamut-mapping algorithm. For illustration of this technique we can choose the soft gamut-mapping approach. Other parameters may be used depending on the gamut-mapping algorithm. DE2000 formula. for example. that selects an exhaustive set of combinations of b and a within their limits. introduced in Section 7. . a best (that minimizes the merit function) is picked for that cluster. Once optimal mapping parameters are obtained for each candidate. Note that the above calculation can be done for more than one gamut-mapping algorithm for each cluster. A merit function is constructed based on a highlevel gamut-mapping strategy with each cluster.6. a*.. The merit-based gamut mapping uses a closed-loop optimization technique (Figure 7. or optimum.62.g.

63 (See color insert following page 428. it is not impractical to imagine more than 50 clusters being processed intelligently by the computer with 90 different gamut-mapping algorithms.398 Control of Color Imaging Systems: Analysis and Design Clusters can be formed by running the vector quantization algorithm (Section 2. all details in the shadow regions may be lost. with this method. . top right: without BPC. Therefore.7) for the L*a*b* nodes or can be formed in or around the vicinity of the corners.63). an L* mapping technique for black point compensation similar to the one shown in Figure 7. bottom: with BPC). They often contain information the viewer knows should be there (the folds of a dark coat. For implementing the black point compensation (i.6 BLACK POINT COMPENSATION RGB images often contain colors that are darker than the darkest color a printer can make.e. In the entire color space. though at a reduced contrast.. input L* to output L* map along the neutral axis) FIGURE 7. However.60 is often used for RGB images.) Example of black point compensation (top left: original.6. and faces. and at the cost of lightening some of the darkest colors (see Figure 7. Minimum color error can be obtained by mapping the out-of-gamut colors to the darkest color the printer can make. in that case. It retains the details for dark colors. the strands of dark hair) and their absence can be very disturbing. 7.4. edges.

The reproducible color gamut of the printer represented in the printer model is the volume covered by the second set of colors.. At the same time if the test colors are sufﬁciently dense and well distributed. It is not impossible to ﬁnd that colors that can be produced using one proﬁle may be impossible with another. The gamut volume and round trip accuracy is computed as follows. 0. Ideally.g. type of GCRs..g.0. Otherwise. The output of B2A tag is then fed as inputs to the A2B tag to produce second set of L*a*b* values. After that.7 EVALUATION OF PROFILES 7. Obtain their estimated L*a*b* values for each of the test colors using the A2B tag. Create a large number of in-gamut test colors in CMYK space (e. Hence. and node spacings are some of the factors that contribute to the gamut of colors that a printer can produce. DE2000 or any other DE) values for a set of in-gamut test colors between their input L*a*b* values for the B2A tag and the output L*a*b* values from the A2B tag.0) and ﬁne tune the parameters a and b. such as an in-line or ofﬂine spectrophotometer for a set of CMYK test colors. Colorants of toner (or ink). A2B tag represents the printer model and B2A tag contains the gamut-mapped inverse LUT..g. process physics. The gamut volume can be obtained using the methods described in Ref.1 GAMUT UTILIZATION AND ROUND TRIP ACCURACY Color gamut and its utilization depend on various factors. . First a set of L*a*b* values are measured using a sensor.g.7.. The printing industry has access to numerous commercially available proﬁling packages. 7. 5000). Using them in sequence for a set of in-gamut test colors described in L*a*b* space can sometimes * give a good indication of the quality of inversion. we can also generate the gamut volume.Three-Dimensional Control of Color Management Systems 399 using soft gamut mapping with ray-based control method.0) x0 along the neutral axis (e. chromatic greens=reds) where the deﬁciencies are present.0. then one should produce a vector plot between the two sets of L*a*b* values to identify the regions of colors (e. One way to quantify the gamut utilization is through computational procedures by using A2B and B2A tags of the ICC Proﬁle. If the gamut volume is not 100%.12). it is important to assess the gamut utilization for a given proﬁle from an awareness and competitive point of view. The roundtrip accuracy is equal to the DE between ﬁrst and the second set of L*a*b* values. each node color in the inverted LUT is updated with the new input L* to output L* map (see Problem 7. Now use these estimated L*a*b* values as inputs to the B2A tag. Another way to quantify the gamut utilization is by using the printer instead of the A2B tag of the ICC proﬁle. we can quantify the round trip accuracy. the printer may not reproduce all the colors in an RGB image even though the printer’s CMYK gamut is capable of producing those colors. we deﬁne xc equal to zero (e. 50. the inversion round trip error for a colorimetric proﬁle must be zero and the round trip gamut volume should be close to 100% when compared to the native CMYK gamut. By computing DE (DEab. [135]. The reproducible gamut is generally smaller than the available process gamut. inversion methods.

78 12.65 b* À57. and 12 faces. For determining round trip accuracy. averaging with multiple sets.00 72.34 27.14 CMYK and L*a*b* Values for a Typical EP Production Printer Gamut Corners Corners Cyan Magenta Yellow White Red Green Blue Dark cyan Dark magenta Dark yellow Dark black Dark blue Dark green Dark red C 255 0 0 0 0 255 255 255 0 0 0 255 255 0 M 0 255 0 0 255 0 255 0 255 0 0 255 0 255 Y 0 0 255 0 255 255 0 0 0 255 0 0 255 255 K 0 0 0 0 0 0 0 255 255 255 255 255 255 255 L* 50. Measurement procedures adopted during the printer characterization should match those during the proﬁle validation stage.63 0.02 16.80 35. 7. Figures 7.07 À21.67 .09 À51.. same stock backing.34 À63.41 a* À29.64a and b show these corners when viewed from the top and bottom.34 20. use the measured L*a*b* values as inputs to the B2A tag. Create patches for the output of B2A tag and print and measure.89 17. A vector plot of these L*a*b* values will give additional information about the rendering accuracy. Figure 7. have to be strictly followed when a spectrophotometer is used. Otherwise.90 À7.64c shows all 14 lines emanating through the corners until they touch the sides of the L*a*b* cube. which is particularly useful to accessing the spot roundtrip inaccuracy in localized regions of the color space.01 23.70 27. then create patches using the measured L*a*b* values and print them through the ICC proﬁle and measure the printed patches.80 1.00 51.68 À10. same white point.51 17.79 À6.14. validation will not be accurate.21 100.01 47.33 17.71 À3. Also obtain the round trip accuracy statistics.400 Control of Color Imaging Systems: Analysis and Design Measurement guidelines.11 1. A generic CMYK printer gamut has at least 14 distinct corners. They can be used to identify unsmooth color gradients and help ensure smoothness in CMYK formulations in critical regions of the transformation. This procedure will eventually result in a second set of L*a*b* values.17 30.05 50.12 À0.2 GAMUT CORNER PLOTS AND NEUTRAL RESPONSE Gamut corner plots can give additional measurement of the quality of inversion and gamut-mapping algorithms.70 27.65(a to d) and 7.17 À33. 24 edges.7.00 À24.40 À0.98 84.47 111. etc.37 43.08 24.93 50.83 95. and Figure 7. The corner CMYK and corresponding L*a*b* values for a representative printer are shown in Table 7. such as same paper. If this procedure is difﬁcult to implement.66 TABLE 7. Compare the gamuts between the second and the ﬁrst set of L*a*b* values. respectively.11 21.64 14.

show the CMYK response (left) and round trip accuracy (right) when a line stimulus is used from L* ¼ 100. The line can be expressed in L*a*b* space by the line equation with a start point {L0 a0 b0} and . and blue corners from 100.0 point).) (a) (left) and (b) (right): Gamut views from top and from bottom. representing the points on the gamut corner axis. is extended until it touches the sides of the L*a*b* cube.0 point). white. a* ¼ b* ¼ 0 in Figure 7. The space between 0 and 100 units is the in-gamut axis and the space between the red vertical lines (at 100) to the end is the out-of-gamut region. magenta.Three-Dimensional Control of Color Management Systems 401 FIGURE 7. The corner point shown is always at 100. yellow. respectively. Similar plots are shown from L* ¼ 50. white. The x-axis. magenta. The corner is shown with a thick vertical line in each of these plots. a* ¼ b* ¼ 0) to the corner. (d) lines to gamut corners (cyan.0.0. a* ¼ b* ¼ 0 point to 14 different gamut corners. yellow.64 (See color insert following page 428. The x-axis contains 101 discrete points along the line between the starting point (L* ¼ 50. (c) lines to gamut corners (cyan.65(e–i). and blue corners from 50.

66).0 point (‘0’ on x-axis corresponds to 100. Note that when the line stimulus is given between {100. in Equation 7.0}. yellow corners from 100. the round trip accuracy corner plots from (50.95.95 and 7.0.0) show lower accuracy numbers as compared with similar corner plots from (100. magenta.0.402 Control of Color Imaging Systems: Analysis and Design corner point {Lc ac bc} shown by Equations 7. the corner plot shows the neutral response of the inversion algorithm (Figure 7. the line will intersect the gamut corner point {Lc ac bc}.0) point.0. 250 200 CMYK values Corners = 255 0 0 0 ΔE2000 values (roundtrip) 8 7 6 5 4 3 2 1 0 0 Corners = 255 0 0 0 K C M Y 150 100 50 0 0 20 40 60 80 100 120 140 160 180 From 0 to gamut corner Corners = 0 255 0 0 ΔE2000 values (roundtrip) 20 40 60 80 100 120 140 160 180 From 0 to gamut corner Corners = 0 255 0 0 250 200 CMYK values 8 7 6 5 4 3 2 1 0 0 150 100 50 0 K C M Y 0 50 100 From 0 to gamut corner Corners = 0 0 255 0 150 50 100 From 0 to gamut corner Corners = 0 0 255 0 150 250 200 CMYK values 8 ΔE2000 values (roundtrip) 150 100 50 0 K C M Y 7 6 5 4 3 2 1 0 0 20 40 60 80 From 0 to gamut corner 100 0 20 (a) 40 60 80 From 0 to gamut corner 100 FIGURE 7.65 (a) CMYK response (left) and corresponding round trip accuracy (right) to cyan.96.0} and the dark corner {0. . when i ¼ 0 the line will start at x ¼ x0.0. Clearly.0 point). When i ¼ N. For the purpose of drawing corner plots.0.0.

0 point). blue corners from 100.0.65 (continued) (b) CMYK response (left) and corresponding round trip accuracy (right) to red.0 point (‘0’ on x-axis coresponds to 100. green.0. Clearly. If the CMYK response does not show a smooth continuous function with respect to x-axis. then we can expect 250 200 CMYK values 150 100 50 0 0 Corners = 0 255 255 0 ΔE2000 values (roundtrip) K C M Y 8 7 6 5 4 3 2 1 0 0 Corners = 0 255 255 0 20 40 60 80 100 120 140 160 180 From 0 to gamut corner Corners = 255 0 255 0 ΔE2000 values (roundtrip) K C M Y 8 7 6 5 4 3 2 1 0 20 40 60 80 100 120 140 160 180 From 0 to gamut corner Corners = 255 0 255 0 250 200 CMYK values 150 100 50 0 0 20 40 60 80 100 120 140 160 From 0 to gamut corner Corners = 255 255 0 0 ΔE2000 values (roundtrip) K C M Y 0 20 40 60 80 100 120 140 160 From 0 to gamut corner Corners = 255 255 0 0 250 200 CMYK values 150 100 50 0 8 7 6 5 4 3 2 1 0 0 (b) 0 20 40 60 80 100 From 0 to gamut corner 120 20 40 60 80 100 From 0 to gamut corner 120 FIGURE 7. the corner plots provide useful indication of the round trip accuracy in critical regions. (continued ) .Three-Dimensional Control of Color Management Systems 403 We deliberately chose the proﬁles with comparatively higher round trip accuracy for near-boundary nodes to highlight the usefulness of these plots in identifying DE2000 errors. They also provide a visual indication of the CMYK response to continuous L*a*b* stimuli along the line.

which can lead to contours. dark yellow corners from 100. In such cases. smoothing methods should be applied to smooth out the transitions..65 (continued) (c) CMYK response (left) and corresponding round trip accuracy (right) to dark cyan. dark magenta.0.65c and d. Sometimes. In Figure 7. tints).0 print (‘0’ on x-axis corresponds to 100. highly dense LUTs can also be built to reduce such jumps in CMYK formulations and improve roundtrip accuracy.0. higher LUT size 250 200 CMYK values 150 100 50 0 Corners = 255 0 0 255 ΔE2000 values (roundtrip) K C M Y 8 7 6 5 4 3 2 1 0 Corners = 255 0 0 255 0 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner 250 200 CMYK values 150 100 50 0 K C M Y Corners = 0 255 0 255 ΔE2000 values (roundtrip) 8 7 6 5 4 3 2 1 0 0 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner Corners = 255 0 0 255 0 20 40 60 80 From 0 to gamut corner Corners = 0 0 255 255 100 0 20 40 60 80 From 0 to gamut corner Corners = 0 0 255 255 100 250 200 CMYK values 150 100 50 0 K C M Y 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 20 0 20 (c) 40 60 80 100 From 0 to gamut corner 40 60 80 100 From 0 to gamut corner FIGURE 7.g.404 Control of Color Imaging Systems: Analysis and Design undesirable contours when rendering color sweeps (e. examples of sharp transitions in the round trip accuracy for dark magenta corner [0 255 0 255] and dark blue corner [255 255 0 255] can be seen. . these transitions are caused by sparse nodes in the L*a*b* to CMYK LUT. But. But.0 point).

hence. (continued ) .0.0. may not be suitable for general use. It is also critical to note that the CMYK response of out-of-gamut colors to the line stimuli should also be smooth. and. dark green. and dark red corners from 100.Three-Dimensional Control of Color Management Systems 405 requires more memory. more time for creating proﬁles. If discontinuities occur for the out-of-gamut colors.65 (continued) (d) CMYK response (left) and corresponding round trip accuracy (right) to dark blue.0 point). then different mapping strategies may have to be used to preserve the 250 200 CMYK values 150 100 50 0 Corners = 255 255 0 255 ΔE2000 values (roundtrip) K C M Y 8 7 6 5 4 3 2 1 0 Corners = 255 255 0 255 0 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner Corners = 255 0 255 255 ΔE2000 values (roundtrip) K C M Y 0 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner Corners = 255 0 255 255 250 200 CMYK values 150 100 50 0 0 8 7 6 5 4 3 2 1 0 0 20 20 40 60 80 From 0 to gamut corner Corners = 0 255 255 255 100 40 60 80 From 0 to gamut corner Corners = 0 255 255 255 100 250 200 CMYK values 150 100 50 0 K C M Y 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 20 0 20 (d) 40 60 80 From 0 to gamut corner 100 40 60 80 From 0 to gamut corner 100 FIGURE 7.0 point (‘0’ on x-axis corresponds to 100.

0 point). Otherwise.0.3 VISUAL EVALUATION OF PROFILES No single proﬁle can please everyone.7. Thus. the corner plots typically probe deep into the local regions of a color gamut and can be a good indicator of the local inversion and mapping problems.65 (continued) (e) CMYK resoponse (left) and corresponding round trip accuracy (right) to cyan.0 point (‘0’ on x-axis corresponds to 50. . visual artifacts can be expected from the ﬁnal proﬁle. Visual evaluation is largely inﬂuenced by the observer characteristics. viewing conditions.406 Control of Color Imaging Systems: Analysis and Design smoothness. magenta. and yellow corners from 50. 7.0. and preferences to a certain color and 250 200 CMYK values 150 100 50 0 Corners = 255 0 0 0 ΔE2000 values (roundtrip) 8 7 6 5 4 3 2 1 0 0 Corners = 255 0 0 0 K C M Y 0 20 40 60 80 100 120 140 160 180 200 From 0 to gamut corner Corners = 0 255 0 0 20 40 60 80 100 120 140 160 180 200 From 0 to gamut corner Corners = 0 255 0 0 250 200 CMYK values 150 100 50 0 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 K C M Y 0 50 100 From 0 to gamut corner Corners = 0 0 255 0 150 50 100 From 0 to gamut corner Corners = 0 0 255 0 150 250 200 CMYK values 150 100 50 0 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 20 K C M Y 0 20 40 60 80 100 From 0 to gamut corner (e) 40 60 80 From 0 to gamut corner 100 FIGURE 7.

0.Three-Dimensional Control of Color Management Systems 407 its appearance with the image content. halftone screens. rating experiments are carried out with weights applied to certain image quality attributes.0. (continued ) . green corners from 50. Ultimate color rendering performance on a variety of print media depends on how well the proﬁles perform for different paper stocks. Often. red.0 point (‘0’ on x-axis corresponds to 50.65 (continued) (f) CMYK response (left) and corresponding round trip accuracy (right) to white. and attribute tables are populated by comparing images to proofs.0 point). and print 250 200 CMYK values 150 100 50 0 Corners = 0 0 0 0 ΔE2000 values (roundtrip) 8 7 6 5 4 3 2 1 0 0 Corners = 0 0 0 0 K C M Y 0 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner Corners = 0 255 255 0 ΔE2000 values (roundtrip) 10 20 30 40 50 60 70 80 90 100 From 0 to gamut corner Corners = 0 255 255 0 250 200 CMYK values 150 100 50 0 8 7 6 5 4 3 2 1 0 0 20 K C M Y 0 20 40 60 80 100 120 140 160 180 From 0 to gamut corner Corners = 255 0 255 0 ΔE2000 values (roundtrip) 40 60 80 100 120 140 160 180 From 0 to gamut corner Corners = 255 0 255 0 250 200 CMYK values 150 100 50 0 8 7 6 5 4 3 2 1 0 0 20 K C M Y (f) 0 20 40 60 80 100 120 140 160 From 0 to gamut corner 40 60 80 100 120 140 160 From 0 to gamut corner FIGURE 7.

dark cyan. Many of the proﬁle-induced image quality defects can be spotted easily by performing soft prooﬁng on a calibrated monitor. .65 (continued) (g) CMYK response (left) and corresponding round trip accuracy (right) to blue.408 Control of Color Imaging Systems: Analysis and Design conditions. dark yellow corners from 50.0 point). the blocking in hair can be easily observed. by using Adobe PhotoShop software or MATLAB image processing tool box.0.0.0 point (‘0’ on x-axis corresponds to 50. For example. which is a good example of not having proper black point compensation. 250 200 CMYK values 150 100 50 0 0 K C M Y Corners = 255 255 0 0 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 Corners = 255 255 0 0 50 100 From 0 to gamut corner Corners = 255 0 0 255 150 50 100 From 0 to gamut corner Corners = 255 0 0 255 150 250 200 CMYK values 150 100 50 0 0 K C M Y 8 ΔE2000 values (roundtrip) 100 7 6 5 4 3 2 1 0 0 20 20 40 60 80 From 0 to gamut corner Corners = 0 255 0 255 40 60 80 From 0 to gamut corner Corners = 0 255 0 255 100 250 200 CMYK values 150 100 50 0 0 K C M Y 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 20 20 (g) 40 60 80 100 From 0 to gamut corner 120 40 60 80 100 From 0 to gamut corner 120 FIGURE 7.

and dark blue corners from 50.Three-Dimensional Control of Color Management Systems 409 If the in-gamut colors in the original image do not match the colors in the rendered image. (continued ) .65 (continued) (h) CMYK response (left) and corresponding round trip accuracy (right) to dark yellow. then the errors can be associated to round trip accuracy with the actual printing device.0 point). dark black.0. if a printer model is used as a surrogate characterization of the printer.0 point (‘0’ on x-axis corresponds to 50. 250 200 CMYK values 150 100 50 00 K C M Y Corners = 0 0 255 255 ΔE2000 values (roundtrip) 8 7 6 5 4 3 2 1 00 20 Corners = 0 0 255 255 20 40 60 80 100 120 From 0 to gamut corner Corners = 0 0 0 255 140 40 60 80 100 120 From 0 to gamut corner Corners = 0 0 0 255 140 250 200 CMYK values 150 100 50 0 K C M Y 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 20 0 20 40 60 80 100 From 0 to gamut corner Corners = 255 255 0 255 120 40 60 80 100 From 0 to gamut corner Corners = 255 255 0 255 120 250 200 CMYK values 150 100 50 0 8 ΔE2000 values (roundtrip) 7 6 5 4 3 2 1 0 0 20 K C M Y 0 20 (h) 40 60 80 100 From 0 to gamut corner 40 60 80 100 From 0 to gamut corner FIGURE 7. This could be due to inversion errors or the errors in the printer model.0.

0.0.0 point (‘0’ on x-axis corresponds to 50.66 CMYK response and round trip accuracy to neutral axis—gamut corner plot from 100.0 point to dark black corner. if iterations are carried out directly on the printer and the measurement process and system noise are assumed to be well within the reasonable limits when compared to the total round trip accuracy of the actual printing device. (‘0’ on x-axis corresponds to 100. dark red corners from 50.0 point). Improving accuracy can give rise to . Or else.65 (continued) (i) CMYK response (left) and corresponding round trip accuracy (right) to dark green.0. the error may be only due to the inversion process.410 Control of Color Imaging Systems: Analysis and Design Corners = 255 0 255 255 ΔE2000 values (roundtrip) 250 200 CMYK values 8 7 6 5 4 3 2 1 0 0 20 Corners = 255 0 255 255 150 100 50 0 0 K C M Y 20 40 60 80 100 From 0 to gamut corner Corners = 0 255 255 255 120 40 60 80 100 From 0 to gamut corner Corners = 0 255 255 255 120 250 200 CMYK values 8 ΔE2000 values (roundtrip) 150 100 50 0 K C M Y 7 6 5 4 3 2 1 0 0 20 40 60 80 100 From 0 to gamut corner 120 0 20 (i) 40 60 80 100 From 0 to gamut corner 120 FIGURE 7. 250 200 CMYK values 150 100 50 0 K C M Y Corners = 0 0 0 255 ΔE2000 values (roundtrip) 8 7 6 5 4 3 2 1 0 0 20 40 60 80 From 0 to gamut corner 100 Corners = 0 0 0 255 0 20 40 60 80 From 0 to gamut corner 100 FIGURE 7.0 point).0.

In Figure 7. These colors are inside the printer gamut (see Figure 7. grass contours occurred due to an inaccuracy in the proﬁles created by the MM algorithm. the in-gamut colors are shifted deliberately to offset bias in the printing system or to render more preferred outcome. Nonsmooth color transitions can cause contours.67 (See color insert following page 428. if the CMYK formulations are not smooth between nodes of the proﬁle LUT. improved image details and can be an enabler for photo rendering if the halftone screens and printer resolution is up to the desired quality.67). For example. sky colors are generally centered on a 280 degree hue angle and may be purposely detuned with a hue shift for preference matching during the proﬁle creation stage. which is largely due to unpredictability associated with nearboundary colors. Often. such as in skin tones.Three-Dimensional Control of Color Management Systems 411 FIGURE 7. but has the effect to render more preferred color. we may see gamut-mapping issues like many-to-one mapping. Emulation with respect to industry color standards like GRACoL.65. contouring can occur when color sweeps intersect the gamut boundary. Japan Color or other source color spaces becomes easy provided the colorants have enough gamut coverage. the control-based inversion).39 for performance with MM algorithm vs. ISO. when the tone scales appear dirty or grainy or nonuniform with black toner. In addition to this. reduced contours. For example. then tone scale suppression may be required. balanced neutrals. Also.) Example of in-gamut color sweeps—susceptible to contours (see Figure 7. Since GCR plays an important role in high quality color reproduction using toner or inks. improved highlights. Any intentionally induced color shifts may not give a good match to the original. Hence the task of achieving smoothness without loosing roundtrip accuracy is a real challenge to high quality reproduction with multidimensional LUTs.38 and in the corner plots (roundtrip accuracy) shown in Figure 7. when color sweeps occur from dark out-of gamut color to light in-gamut color as in human arm (Figure 7.39. we can expect to see contours. .68) or sensitive regions. Inaccuracies in the inversion process can also lead to nonideal behavior in gamut-mapped colors as seen in Figure 7.

NTSC RGB. 7. We adopt relative-to-paper colorimetry to incorporate a standard reference media to reduce paper dependency. A 3DLUT is a major component of the destination ICC proﬁle which is a GCRconstrained. We use the footprint shown in Figure 7. compression. gamut-mapped transformation LUT from RGB triplets to device CMYK space. There are numerous choices for this (Adobe RGB. the control-based inversion). the paper white will have the PCS values of {100. The performance is evaluated for a high resolution 333 LUT using the same printer model. For this example to build an ICC proﬁle for a printing device with media relative colorimetric rendering intent. let us restrict the colorimetric options to a 2 degree observer. Apple RGB. That is. Step 1: Choose a working color space for RGB triplets. 0.g.68 (See color insert following page 428.412 Control of Color Imaging Systems: Analysis and Design FIGURE 7. and number of illuminants. and other compromises to the basic colorimetric proﬁle.8 AN EXAMPLE SHOWING HOW TO BUILD MULTIDIMENSIONAL INVERSE LUT The ICC proﬁle architecture assumes a reference color space in L*a*b* or XYZ values (called PCS) for standard observers. CIE RGB. All reﬂectance measurements are done with a speciﬁed measurement geometry (0=45 or 45=0).. perceptual.) can also be easily built by incorporating proper tone scale adjustments.1 to describe the steps. Hence we ﬁrst show the detailed steps involved in building the 3DLUT starting from a printer model. and a D50 illuminant source with the PCS deﬁned by L*a*b* values.40 for performance with MM algorithm vs. Wide . 0}. we use a low resolution LUT of size 63 to show the numerical values. The ICC proﬁles with other rendering intents (e. measurement geometry. saturation etc.) Example of out-of-gamut color sweeps—susceptible to contours (see Figure 7. ColorMatch RGB. For convenience.

Sb Yb Sb Zb Zi ¼ (1 À xi À yi)=yi with index i ¼ r. YW ¼ 1.0–1. then use the XYZ to Lab equation (see Appendix A). Chromatic adaptation algorithms can be used to convert the source color (sRGB with D65 reference white) into the destination color (sRGB with D50 reference white) by a linear transformation (Bradford.0–1. Yi ¼ 1. The RGB values must be linear (i. if the sRGB color space is used as the working space.0).82521. XYZ scaling.0. Sb ] ¼ [XW Xr ZW ]4 Xg Xb 2 Yr Yg Yb 3À1 Zr Zg 5 Zb (7:102) Sg YW The XYZ values in Equation 7. To limit our options. see also Equations 7. For converting XYZ to RGB the M matrix is inverted (i. 255). [Sr 2 Y Z] ¼ [R G B]M (7:101) 3 Sr Yr Sr Zr Sg Yg Sg Zg 5. b and. for converting RGB to XYZ is given by [X Sr Xr where. It is important to note here that the reference whites are consistent with ICC speciﬁcations.96422. If the values are in the range (0. chromaticity coordinates for a ProPhoto RGB system are given by xr ¼ 0:7347.. ZW).. we choose ProPhoto RGB as the color space. sRGB. This choice was made because the ProPhoto RGB space gives reasonable gamut coverage (see Figure 7. For example. the 3 Â 3 conversion matrix.0).28).101 uses g ¼ 1. For a 2-degree observer and a D50 illuminant [136]. then divide the RGB values by 255. To convert the RGB color space to Lab. and (xb. xb ¼ 0:0366. then direct use of sRGB gives a mismatch between ICC proﬁles deﬁned relative to D50 reference white since sRGB is deﬁned relative to D65 light source. we ﬁrst determine the matrix for converting RGB to XYZ. yb ¼ 0:0001. etc.101 are in the nominal range (0.105 and 7.0. xg ¼ 0:1596. YW.). MÀ1).e. From Appendix A (Table A. yg). they should be adopted to a D50 reference white from their native reference white. XW ¼ 0.2). g. (7:103) . M. yr).Three-Dimensional Control of Color Management Systems 413 Gamut RGB. Gamuts produced by the chosen color space should be compared to the gamut of the printer in Lab space prior to selecting the ﬁnal working color space so that all colors that are reproducible by the printer gamut are covered by the working color space. Equation 7. yb) and its reference white (XW. and Von Kries are popular transforms). M ¼ 4 Sg Xg Sb X b Xi ¼ xi=yi. (xg.106 for use with g 6¼ 1. Given the chromaticity coordinates of the RGB system (xr. They also have to be scaled. yr ¼ 0:2653.e. Before they can be used in a D50 environment. ROMMRGB. ZW ¼ 0. which gives non-linear relationship) and in the nominal range (0. yg ¼ 0:8404. which is dependent on the source reference white and the destination reference white.

Convert the RGB colors to XYZ and then to L*a*b* using Equation 7. is obtained as 0:7977 0:2880 M ¼ 4 0:1352 0:7119 0:0314 0:0001 2 3 0 0 5 0:8252 (7:104) Given an RGB color such that its components are in the nominal range [0. the conversion matrix. A low resolution LUT of size 63 was formed to show numerical values by varying the ﬁrst channel R least rapidly and varying third channel B most rapidly between 0 and 255 with a uniformly sampled RGB grid. After creating the RGB grid points. g ¼ 1. 1.0] and its gamma is g and the RGB system is not sRGB: [X where r ¼ Rg . g ¼ Gg .414 Control of Color Imaging Systems: Analysis and Design Substituting these parameters in Equation 7. M.105 and the equations shown in Appendix A for a 2-degree observer and D50 illuminant. g. r.g.0. We used the . In this example. Step 2: Form a uniformly sampled RGB LUT (e. b ¼ Bg For sRGB.g. 173). Using a proper choice of g (e.. Step 3: Separate L*a*b* nodes into in-gamut and out-of-gamut colors. given the working space deﬁnition for RGB color space. we know the transformation matrix. b values can be found from the following equations: ( r¼ ( g¼ ( b¼ R=12:92 ((R þ 0:055)=1:055)2:4 G=12:92 ((G þ 0:055)=1:055)2:4 B=12:92 ((B þ 0:055)=1:055)2:4 R 0:04045 (7:106) Y Z] ¼ [r g b]M (7:105) R > 0:04045 G 0:04045 (7:107) G > 0:04045 B 0:04045 B > 0:04045 Now.. the data is converted to L*a*b* and then separated into two groups: in-gamut and out-of-gamut colors.8 for ProPhoto RGB) we can determine the XYZ values and hence L*a*b* values. according to the uniform sampling approach described above.102. the color index 1 corresponds to ﬁrst grid point with R ¼ 0 ¼ G ¼ B and index 216 corresponds to the last grid point R ¼ G ¼ B ¼ 255 with increments in between.

93 71.46 85.46 À2. In this example.00 46.15 RGB ! L*a*b* and L*a*b* ! CMYK LUT for In-Gamut Node Colors Color Index 44 45 50 51 80 81 86 87 88 122 123 124 128 129 130 170 171 172 173 RGB Values R 51 51 51 51 102 102 102 102 102 153 153 153 153 153 153 204 204 204 204 G 51 51 102 102 51 51 102 102 102 102 102 102 153 153 153 204 204 204 204 B 51 102 51 102 51 102 51 102 153 51 102 153 51 102 153 51 102 153 204 L Node L*a*b* (Out-of-Gamut Colors) * GCR Constrained CMYK C 111 207 221 227 6 101 107 137 155 21 39 61 62 80 96 20 30 38 51 M 80 142 9 14 218 216 88 105 102 160 154 151 46 67 72 14 30 33 32 Y 85 12 246 145 105 4 209 107 27 206 89 4 229 141 73 249 162 94 33 K 182 92 21 17 110 25 55 28 11 18 12 1 20 5 2 14 4 2 3 a * b * 28.17 46. The gamut-mapped L*a*b* and their corresponding CMYK values are shown in Tables 7.52 27.86 À20.46 85.96 À52.92 0.85 48.79 59.38 69.15 for in-gamut colors.00 36.6.2. we used a simple centroid clipping (Section 7.16 and 7.17.31 39.19 À3. We used the K-restricted GCR method during the inversion process.31 50.27 0.1 to separate the colors into two groups.73 0. The out-of-gamut nodes are shown in Tables 7.79 0. all of the out-of-gamut nodes are mapped to the gamut boundary.01 43.12 10. To create an ICC proﬁle. TABLE 7.72 À2.78 50.24 57. Clearly.00 À4. Step 5: Perform constrained GCR inversion for in-gamut colors.76 À8.16 and 7.00 À24.98 À56.16 28. this LUT can be used to ﬁll the CLUT inside the ICC structure.94 50. These colors are the nodes in L*a*b* space at which the device CMYK values are found that populate the multidimensional CLUT of the ICC proﬁle.24 69.38 À3.6.38 85. The RGB and the corresponding L*a*b* node values are shown in Table 7.94 57.87 50.86 36.38 69.07 0.00 . Now rearrange the in-gamut and out-of-gamut RGB to CMYK LUTs to form a single LUT with color index incremented from 1 to 216.17.46 0.51 15.00 À31.Three-Dimensional Control of Color Management Systems 415 ray-based control model described in Section 7.05 31.00 3.94 50.24 0.24 57.46 85. Table 7.00 4.00 À39.57) with ray-based control model.40 À1.00 0.26 30. Figure 7.27 39.15 contains the CMYK values corresponding to each in-gamut node.55 41.00 98. Step 4: Perform gamut mapping for out-of-gamut colors.

50 À46.15 À47.34 67.87 65.58 13.19 À53.45 23.81 À39.92 103.60 30.77 À39.99 À44.34 À37.32 À23.77 Gamut Mapped CMYK C 143 255 255 193 166 152 209 255 255 255 220 186 202 233 233 238 243 244 183 186 189 191 193 197 167 170 173 174 175 175 158 160 163 165 165 165 0 112 173 157 147 M 112 183 210 200 187 179 1 0 11 56 107 116 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 255 255 231 202 183 Y 120 0 0 0 0 0 255 175 67 9 0 0 255 224 161 106 59 14 238 206 167 130 102 75 222 199 172 146 119 101 216 194 173 153 132 114 178 0 0 0 0 K 255 158 13 0 0 0 113 119 110 62 0 0 15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 163 112 0 0 0 .88 À63.61 25.24 À42.79 À45.69 À185.48 À35.80 À119.416 Control of Color Imaging Systems: Analysis and Design TABLE 7.45 60.73 À53.60 74.97 À49.29 À33.26 60.24 60.30 À161.79 À75.48 À96.98 21.88 À42.41 À38.37 À153.81 À19.91 À131.27 À10.78 27.00 0.84 35.48 67.92 34.86 3.38 À60.58 À3.73 À33.47 À48.52 64.37 52.39 À130.25 40.28 40.24 74.11 À96.80 64.55 À87.65 À21.14 À71.39 44.01 65.77 43.53 À137.97 24.05 31.91 À45.85 20.40 À8.58 87.81 À36.99 À63.37 43.44 23.44 60.02 23.29 À70.36 À44.58 87.19 À106.79 9.27 48.54 À123.37 17.96 75.26 À35.96 À139.98 45.42 À79.23 60.30 À17.21 À42.58 41.19 À42.70 À25.90 À178.85 À54.26 1.91 À168.26 À107.41 À135.44 64.21 67.87 À96.76 À40.81 31.20 a* 0.71 39.96 À42.00 0.75 49.70 4.78 43.95 À47.14 14.88 À124.13 20.36 70.24 60.53 21.01 52.16 31.29 À105.45 60.93 À173.17 13.53 35.77 43.58 87.86 À158.81 À39.29 À146.60 À39.56 46.28 À37.00 À48.08 66.00 À18.77 43.98 À47.85 À65.00 49.68 10.34 À47.80 À3.63 À186.66 À115.69 4.26 36.99 55.62 80.60 74.17 13.41 11.46 60.51 34.41 45.66 64.76 À43.73 À15.43 23.42 77.47 26.49 À68.05 0.78 43.62 44.60 74.18 À38.77 Gamut Mapped L*a*b* a* 1.55 128.24 60.75 À5.79 43.58 87.08 À5.88 À163.08 23.23 60.98 33.38 3.29 16.59 L* 10.18 24.03 0.46 60.15 À41.56 À147.99 102.78 41.74 À104.27 50.07 67.26 À30.07 À62.94 30.43 63.96 À86.44 À52.88 18.16 6.60 87.56 À46.60 74.78 60.16 RGB ! L*a*b* ! CMYK LUT for Out-of-Gamut Node Colors (between Color Index 1–108) Color Index 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 RGB Values R 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 51 51 51 51 51 G 0 0 0 0 0 0 51 51 51 51 51 51 102 102 102 102 102 102 153 153 153 153 153 153 204 204 204 204 204 204 255 255 255 255 255 255 0 0 0 0 0 B 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 Node L*a*b* (Out-of-gamut Colors) L* 0.59 À182.19 13.67 29.00 6.60 74.82 À40.02 À15.63 À45.32 À40.01 0.44 23.46 43.91 À12.57 1.94 À22.75 68.58 87.83 b* 2.55 25.40 90.92 50.85 55.45 23.23 29.60 14.79 150.18 13.06 40.94 67.41 22.73 À43.10 45.26 À172.28 60.45 32.01 38.51 À39.75 49.63 b* 0.53 À32.42 À49.

19 46.59 75.67 80.72 Gamut Mapped L*a*b* a* 30.55 b* À149.68 13.14 61.89 40.78 À108.98 18.46 À38.56 À56.67 15.96 35.51 10.33 54.16 28.34 88.62 À41.22 Gamut Mapped CMYK C 139 45 208 177 163 163 255 255 201 177 183 186 188 192 195 161 165 168 171 170 171 154 157 159 161 162 162 0 8 95 103 108 109 1 123 123 121 56 154 143 M 172 51 156 147 141 23 5 10 67 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 255 255 232 195 176 159 188 184 158 150 52 104 105 Y K 0 0 255 180 0 0 0 0 0 0 255 53 69 13 11 1 0 0 255 0 214 0 170 0 120 0 89 0 52 0 229 0 203 0 172 0 143 0 114 0 95 0 219 0 198 0 174 0 152 0 129 0 111 0 215 72 38 78 0 0 0 0 0 0 0 0 255 103 0 0 0 0 0 0 255 107 0 0 0 0 (continued) .62 b* À34.69 À123.97 À20.01 46.60 75.58 À0.97 À98.00 46.62 35.01 61.23 47.70 55.88 À0.35 46.80 À38.69 À148.31 81.87 36.91 À16.17 53.64 47.72 37.77 38.76 À164.43 À32.26 À48.23 68.55 47.08 106.04 À98.14 130.19 À70.08 97.59 75.50 10.82 79.65 À31.26 2.17 28.87 36.28 9.03 39.70 18.60 75.94 À39.03 20.59 48.60 88.97 50.92 À153.77 42.09 À41.33 À101.26 58.18 28.39 À43.30 À68.95 a* 87.44 44.51 À84.15 7.70 À32.05 À38.22 28.17 À58.86 À40.29 50.74 À135.44 92.08 À23.07 61.14 60.76 35.64 75.83 À59.51 37.09 48.58 À50.23 36.79 87.34 88.34 À99.63 61.28 44.11 61.01 À42.31 103.63 61.01 46.65 À91.24 À157.06 31.35 28.67 À46.48 55.92 À38.41 47.62 65.18 15.49 À161.70 4.86 36.08 12.82 À85.40 À163.07 67.82 54.57 64.08 À39.16 61.62 48.13 27.35 À41.22 28.70 À38.72 À17.15 65.21 28.63 À38.27 À7.68 À127.00 66.21 28.65 38.04 À46.95 68.16 (continued) RGB ! L*a*b* ! CMYK LUT for Out-of-Gamut Node Colors (between Color Index 1–108) Color Index 42 43 46 47 48 49 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 82 83 84 85 89 90 RGB Values R 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 51 102 102 102 102 102 102 102 102 102 102 102 102 102 G 0 51 51 51 51 102 102 102 102 153 153 153 153 153 153 204 204 204 204 204 204 255 255 255 255 255 255 0 0 0 0 0 0 51 51 51 51 102 102 102 B 255 0 153 204 255 0 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 153 204 255 0 204 255 Node L*a*b* (Out-of-gamut Colors) L* 13.26 À37.46 36.29 50.07 152.41 À36.83 3.74 63.52 À38.26 57.34 88.94 À15.03 48.62 36.34 88.54 12.02 À93.80 À32.94 49.64 61.27 À44.14 À47.Three-Dimensional Control of Color Management Systems 417 TABLE 7.08 À45.55 À43.80 À30.31 À40.32 83.56 À71.34 88.13 52.61 37.96 52.05 À83.92 66.48 32.67 67.25 12.01 62.12 10.95 35.64 0.22 28.16 À56.50 À28.39 À131.94 50.58 30.71 À122.93 À30.57 L* 46.77 65.76 À78.08 À66.37 68.77 64.24 42.49 À134.40 65.21 28.77 À123.59 75.01 À43.83 87.97 À96.85 À2.15 À29.49 À30.99 À38.27 À39.13 26.45 51.67 À116.75 64.64 58.43 À13.89 0.63 61.95 50.12 7.15 61.06 À39.09 68.10 À3.33 81.66 37.88 50.39 À40.93 44.63 61.11 23.68 51.82 À36.16 3.14 61.54 À32.

89 63.90 64.20 46.418 Control of Color Imaging Systems: Analysis and Design TABLE 7.15 46.68 82.10 173 16 0 0 À33.71 62.65 31.69 64.61 146 0 204 0 À36.76 152 0 102 0 TABLE 7.18 2.51 106.36 b* 69.07 80.77 70.39 40.43 103.45 87.16 (continued) RGB ! L*a*b* ! CMYK LUT for Out-of-Gamut Node Colors (between Color Index 1–108) Color Index 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 RGB Values R 102 102 102 102 102 102 102 102 102 102 102 102 102 102 102 102 102 102 G 153 153 153 153 153 153 204 204 204 204 204 204 255 255 255 255 255 255 B 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 Node L*a*b* (Out-of-gamut Colors) L* 64.68 35.19 90.89 8.98 77.07 47.41 À111.86 153 0 151 0 11.98 77.17 RGB ! L*a*b* ! CMYK LUT for Out-of-Gamut Node Colors (between Color Index 109–216) Color Index 109 110 111 112 113 114 115 116 117 118 119 RGB Values R 153 153 153 153 153 153 153 153 153 153 153 G 0 0 0 0 0 0 51 51 51 51 51 B 0 51 102 153 204 255 0 51 102 153 204 Node L*a*b* (Out-of-Gamut Colors) L* 40.39 40.18 À39.65 À121.74 53.65 À19.45 À60.35 À43.95 159 0 136 0 À35.12 69.05 76.53 À24.90 64.90 L* 62.45 85.93 À123.22 151 0 212 0 À37.11 À75.46 Gamut Mapped CMYK C 0 0 26 60 70 78 0 0 14 61 73 M 251 247 233 204 177 156 224 237 234 197 159 Y 221 95 0 0 0 0 255 128 0 0 0 K 0 0 0 0 0 0 9 0 0 0 0 .51 70.38 75.12 69.16 À16.28 51.84 À88.98 64.20 46.90 42.98 90.31 55.01 À118.91 160 0 74 0 À33.66 70.51 68.63 b* 111.90 64.71 57.34 71.09 À8.90 64.72 À35.18 À86.39 40.90 À37.92 À88.39 40.75 47.22 À25.97 124 18 255 30 À39.55 50.37 80.39 À46.45 À34.89 46.64 23.59 106.48 À34.20 90.20 90.51 46.45 63.19 90.54 29.64 64.20 46.48 44.05 104.83 67.52 53.11 69.40 46.62 14.90 64.32 À122.98 77.87 48.83 113.98 À77.39 43.38 61.96 À67.57 74.33 À6.16 b* 50.71 À102.29 52.32 À39.34 24.66 57.94 20.56 À115.06 À23.85 69.81 Gamut Mapped L*a*b* a* b* Gamut Mapped CMYK C M Y K À22.68 149 0 252 0 À36.98 77.88 0.60 64.80 À8.20 a* À49.91 77.59 13.39 40.07 À18.19 90.81 70.83 À12.25 143 0 229 0 À35.76 79.95 67.36 46.37 158 0 102 0 À33.51 134.68 28.86 64.34 À83.40 153 0 121 0 À35.20 a* 102.63 21.96 155.39 150 0 176 0 À35.22 70.75 À6.83 À11.69 Gamut Mapped L*a*b* L* 47.35 180 0 38 0 À24.19 84.90 a* 69.03 156 0 176 0 À37.61 À80.98 77.48 À42.13 31.52 54.20 À49.05 165 0 244 0 À42.98 68.61 À30.47 176 0 99 0 À34.87 109.19 64.58 À19.07 À7.22 À48.83 49.13 18.20 46.30 173 0 177 0 À40.

73 b* À26.12 75.08 74.36 92.67 À59.07 39.64 75.29 40.81 L* 55.24 69.07 À32.13 À19.28 55.67 57.38 81.79 28.13 64.03 À77.04 55.78 À61.34 À37.71 56.56 41.29 À40.12 À31.61 À71.02 65.36 81.02 2.23 74.38 69.19 66.63 62.00 À27.01 51.62 48.21 105.47 59.07 77.76 À30.16 À7.36 81.29 64.52 À29.48 76.42 51.08 À29.99 71.18 63.13 64.79 140.73 108.92 24.98 62.15 42.75 98.17 (continued) RGB ! L*a*b* ! CMYK LUT for Out-of-Gamut Node Colors (between Color Index 109–216) Color Index 120 121 125 126 127 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 RGB Values R 153 153 153 153 153 153 153 153 153 153 153 153 153 153 153 153 153 153 153 204 204 204 204 204 204 204 204 204 204 204 204 204 204 204 204 204 204 204 204 204 G 51 102 102 102 153 153 153 204 204 204 204 204 204 255 255 255 255 255 255 0 0 0 0 0 0 51 51 51 51 51 51 102 102 102 102 102 102 153 153 153 B 255 0 204 255 0 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 Node L*a*b* (Out-of-Gamut Colors) L* 46.28 75.21 57.86 25.45 112.59 À42.08 110.02 57.68 5.Three-Dimensional Control of Color Management Systems 419 TABLE 7.73 122.56 62.84 À16.81 76.45 31.92 À23.58 Gamut Mapped L*a*b* a* 37.76 73.20 19.95 55.89 À2.75 76.28 75.33 66.45 95.59 53.02 55.69 36.13 64.15 58.16 68.18 58.13 64.34 58.79 À29.65 48.17 73.53 À30.57 105.81 57.39 0.99 À30.98 48.96 1.59 À27.18 54.45 20.01 129.36 81.38 69.62 À20.06 70.79 34.01 51.58 52.92 71.94 6.91 69.07 45.90 À44.33 58.79 Gamut Mapped CMYK C 83 14 86 96 45 105 111 94 115 117 124 126 129 120 120 124 129 131 132 0 0 0 0 37 54 0 0 0 0 35 54 0 0 0 0 31 56 6 0 0 M 148 121 130 123 43 66 82 8 0 0 0 0 0 0 0 0 0 0 0 206 196 202 199 181 159 197 190 190 190 171 152 158 176 155 155 149 132 98 111 115 Y 0 255 0 0 255 0 0 255 231 178 118 74 29 247 210 177 144 110 87 189 113 66 3 0 0 207 127 76 14 0 0 237 173 96 35 0 0 255 216 134 K 0 50 0 0 61 0 0 23 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 18 0 0 (continued) .28 55.36 81.61 a* 90.72 52.29 87.13 74.72 119.87 92.73 À9.87 51.26 69.53 À20.21 72.61 74.81 À9.72 À16.70 77.70 62.56 44.82 62.63 62.81 À34.62 58.97 À74.36 81.98 44.97 54.39 À84.72 75.08 29.18 2.05 À36.22 71.36 33.87 92.95 39.87 79.26 124.77 À80.43 10.54 58.62 84.01 51.61 74.82 b* À92.00 49.87 À26.27 58.94 55.50 À80.96 46.85 38.66 À73.28 91.76 121.06 50.23 72.13 64.38 66.87 92.15 74.86 45.77 À12.64 80.51 12.12 111.46 38.65 106.73 À29.83 18.34 À29.68 62.72 0.35 110.01 51.62 63.38 À4.49 À24.00 22.89 127.73 À52.12 À43.35 74.66 À7.17 À1.19 69.17 123.68 6.68 À48.85 À16.16 À19.29 55.61 À27.01 À78.13 160.28 55.24 57.38 33.35 23.92 29.28 55.65 59.87 92.44 12.84 128.33 À20.24 57.11 À27.17 30.02 51.30 À52.28 56.29 55.63 26.55 À27.56 72.87 92.74 72.76 7.52 56.79 À76.31 46.30 À26.08 40.

71 52.46 64.95 64.57 104.15 À24.85 À14.34 23.25 126.27 À22.12 À42.18 31.38 83.00 72.75 Gamut Mapped CMYK C 0 28 60 29 68 73 82 83 85 89 94 0 0 0 0 0 24 0 0 0 0 0 21 0 0 0 0 0 17 0 0 0 0 0 11 0 0 0 0 0 M 105 105 103 35 40 3 0 0 0 0 0 187 176 166 175 174 160 180 173 158 158 161 157 156 160 150 136 148 145 122 127 129 117 114 117 85 89 89 93 84 Y K 59 0 0 0 0 0 255 35 0 0 255 15 221 0 178 0 124 0 91 0 54 0 177 0 119 0 85 0 50 0 15 0 0 0 184 0 128 0 90 0 54 0 20 0 0 0 205 0 154 0 103 0 65 0 28 0 0 0 225 0 176 0 123 0 81 0 39 0 0 0 255 0 203 0 142 0 101 0 50 0 .84 117.89 22.36 2.58 33.82 61.21 34.60 60.81 À8.10 b* 9.24 22.31 16.82 67.08 20.23 18.02 65.28 78.25 61.20 71.90 110.59 18.22 38.07 90.15 À14.43 47.88 45.54 98.73 61.64 138.61 60.79 64.44 67.06 90.44 À20.10 À62.62 85.420 Control of Color Imaging Systems: Analysis and Design TABLE 7.61 60.23 77.06 9.85 61.73 83.15 70.42 89.93 33.60 61.29 80.14 53.77 61.71 À43.06 90.41 128.37 65.35 40.61 À38.19 74.69 100.94 63.35 À25.60 60.38 60.86 81.81 27.83 b* 9.88 88.75 49.42 0.93 À19.52 42.68 64.90 1.36 À34.11 16.92 À33.66 130.53 54.20 71.29 80.19 96.63 30.29 80.02 À18.69 22.01 43.55 À39.10 42.49 29.94 63.65 69.84 78.77 49.11 19.08 À3.61 60.25 70.35 38.89 65.94 L* 73.98 155.46 35.60 60.37 À37.84 83.19 96.61 50.81 À8.97 À14.06 À21.22 145.05 53.89 58.67 7.17 (continued) RGB ! L*a*b* ! CMYK LUT for Out-of-Gamut Node Colors (between Color Index 109–216) Color Index 166 167 168 169 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 RGB Values R 204 204 204 204 204 204 204 204 204 204 204 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 G 153 153 153 204 204 255 255 255 255 255 255 0 0 0 0 0 0 51 51 51 51 51 51 102 102 102 102 102 102 153 153 153 153 153 153 204 204 204 204 204 B 153 204 255 0 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 255 0 51 102 153 204 Node L*a*b* (Out-of-Gamut Colors) L* 74.50 55.62 77.29 15.14 30.73 74.61 0.65 80.59 143.19 À36.68 À6.33 36.19 96.18 70.85 À67.59 26.21 80.21 69.16 À16.59 70.95 63.18 3.44 20.29 48.95 63.95 71.25 77.84 46.95 63.21 71.77 147.06 165.32 À40.51 4.29 80.77 74.46 85.74 71.58 127.21 126.81 49.32 93.85 70.54 41.34 34.29 80.15 122.49 À13.28 132.72 67.75 Gamut Mapped L*a*b* a* 32.76 À4.31 29.42 55.07 a* 36.62 87.70 65.50 140.97 À22.62 63.46 14.94 3.84 64.07 90.68 34.25 98.46 49.17 139.16 79.39 56.29 90.92 102.52 104.61 74.76 À12.58 À49.28 79.18 33.61 74.78 2.47 96.94 61.33 37.51 68.07 35.34 141.60 60.32 9.13 139.41 31.74 55.60 60.70 51.67 À15.42 À9.49 À6.93 À40.21 71.21 71.19 96.95 À17.87 99.19 60.78 44.02 68.32 À19.01 46.86 50.09 53.94 16.42 82.72 27.19 96.28 106.61 60.61 60.06 44.

00 a* 34. Calculate the modeling errors with respect to the virtual printer model shown in Chapter 10. Use the K-means algorithm to cluster the color data in spectral space.38 69. Determine the parameter matrix using least squares.79 0.00 100.00 100. obtain a quadratic and cubic model for the printer.00 Gamut Mapped CMYK C 0 14 0 0 0 0 1 M 83 27 16 22 20 13 1 Y 0 255 231 171 111 64 0 K 0 20 0 0 0 0 0 PROBLEMS 7. How many basis vectors are needed to obtain a reasonably accurate model? 7.99 0.85 84.48 90.79 25.3 Using the RLS algorithm.6 A clustered PCA-based model is to be constructed for a 104 spectral characterization data.93 0.Three-Dimensional Control of Color Management Systems 421 TABLE 7.50 96.95 À5.07 100.58 52. K ¼ 0]. 7.67 81.27 À1.00 100.00 a* 26.17 (continued) RGB ! L*a*b* ! CMYK LUT for Out-of-Gamut Node Colors (between Color Index 109–216) Color Index 210 211 212 213 214 215 216 RGB Values R 255 255 255 255 255 255 255 G 204 255 255 255 255 255 255 B 255 0 51 102 153 204 255 Node L*a*b* (Out-of-Gamut Colors) L* 90.12 172. obtain a global linear model for a CMY gray color [C ¼ 127 ¼ M ¼ Y. The solution should include some details of the experimental procedure.00 b* À8. 7.93 91.26 100. Plot the model accuracy for the test colors as a function of the number of basis vectors.64 À0.22 À2. Model each cluster with the RLS algorithm. Show the DE (DEab.2 Using the RLS algorithm.00 100.88 0.00 Gamut Mapped L*a*b* L* 81. obtain a local linear model for a CMY gray color [C ¼ 127 ¼ M ¼ Y.1 Using the RLS algorithm. Obtain PCA vectors for each cluster. .41 À3.57 46.5 Use bootstrapping techniques to gain conﬁdence in the PCA-based modeling. K ¼ 0].00 100.16 À4. Use PCA mathematical techniques to analyze the set of multivariate input–output characterization * data.54 98. Plot the performance of your model for test data as a function of the number of clusters.09 À1. 7. DE2000 ) error plot with respect to number of principal components. Is a quadratic and cubic model a good choice for this virtual printer? 7.24 23.41 123.45 À2.4 In a printing system CMYK values are sampled between 0 and 255 in a uniform grid to obtain up to 104 input–output characterization data.47 À5.06 0.00 b* À17. Patches are printed and measured by the spectrophotometer. The solution should include some details of the experimental procedure. The spectra are measured at 31 wavelengths from 400 to 700 nm at 10 nm intervals.72 93.51 97.

bc ¼ 30].7 Simulate a node color convergence near the gamut boundary using a pole* placement algorithm. Image technology colour management—Architecture. Launsby.10 Run a 4 to 3 control loop for the nodes in Problem 7. CA. 2nd edn.90. a0 ¼ 0. Extend Problem 7.13 Repeat the proﬁling example shown in Section 7. P. AIR Academy Press & Associates.9.0). PTR Prentice Hall. in steps of 10 with a* ¼ b* ¼ 0) along the neutral axis. Modeling and identiﬁcation with orthogonal basis functions. Simulate the performance of some test images. Show * DEab convergence errors with and without the best actuator algorithm.2. design a black point compensation method that can retain shadow details. Use the Neugebauer model to represent the printer. Dec. Choose any suitable GCR as nominal CMYK values for the node color.8 for additional in-gamut nodes (say. 1999. Understanding Industrial Designed Experiments.G.11 Produce a ray-based control model for mapping an out-of-gamut spot color described with.8 to build a 173 destination ICC proﬁle using MATLAB. . 2005.6. L. 4.8 7. Using the algorithm described in Section 7. Colorado Springs. Repeat this for another ray drawn toward the same gamut centroid for another suitable out-of-gamut node color. Show DEab convergence errors with and without the best actuator algorithm. proﬁle format. Obtain Gamut corner plots and round trip accuracy (overall and to corner points). 1997. and B. Use the Neugebauer model to represent the printer. If the nodes are not in-gamut change the L* values so that they are inside the printer’s gamut. Upper Saddle River. REFERENCES 1. 36th IEEE CDC Preconference Workshop# 7. 7. 7. Change nominal CMYK values and simulate the converge performance.0. S. ISBN 1-880156-03-2. Van den Hof. What do you see? 7.422 Control of Color Imaging Systems: Analysis and Design 7.90 and simulate converge performance with 3-to-3 control loop.3 or a nonlinear input L* to output L* mapping technique. x0 ¼ [L0 ¼ 50. NJ. Schmidt and R.. and data structure. Ljung. xc ¼ [Lc ¼ 80. Is the convergence error near zero at steady state? Plot actual K separation values along the neutral axis from 4 to 3 control loop and superimpose the values on those obtained from Equation 7. T. 7. B. 2. What do you see? Explain. Ninness. San Diego.9 7. P. Simulate a node color convergence using a pole placement algorithm and a 4 to 3 control-based inversion. seven gray targets with L* of 25–90. Oliveira e Silva.12 Black point compensation algorithm is used to retain shadow details in images since many images contain colors that are darker than the darkest color a printer can make. 3. CO. International Color Consortium Speciﬁcation. ICC. and the gamut centroid. b0 ¼ 0]. ac ¼ 50. Do you still see near zero convergence error? Restrict the K separation using K-restricted function of Equation 7. System Identiﬁcation—Theory for the User.R. Simulate the effects of parameters on shadow details. 1:2004-10 (Proﬁle version 4. Heuberger. Wahlberg.

518. I.595. Oct. 2.K.397. 20. 7. NIP20: Proceedings of the IS&T’s International Conference on Digital Printing Technologies.860. P. Love.435. 11. Jung. Ali.A. 6.332. Hallikainen. 2003. Systems and Applications. Jackson. 19. Dianat.404.809.107. 1990.A.J. J. Sharma. 5. 6. L. US Patent Application. Jung. Characteristic spectra of Munsell colors. Vector-subspace model for color representation. An image output color management system and method.583. Other US Patents 7. 5.902 (Apple). 26. Jaaskelainen. Berns. US Patent 7.272. 6.D. 23.383.147. and T. J. Robbins. L. W. Chiu.262.574. J. Aug.G. 2007. Jolliffe. Diehl.454. UT. and G. Reference database and method for determining spectra using measurements from an LED color sensor. A. Method and apparatus for calculating image correction data and projection system. 2001. Spectral color reproduction with six color output.844. R. P.500. Apr. S. 2006. 2. and T. 21. Dalal. Thieret.770.C. Taplin. 5. and T.N. and K. Dianat.K. Schmitt. D. US Patent 5. Cargill.020. 2003.170.261. in IS&T and SID’s 5th Color Imaging Conference: Color Science. Koditschek. and S. 15. 2004.K.960 (Kodak). US Patent 6.400. J.J. 6. Oct.538.-J. Dec. 24. 10.E. 5. Parkkinen. E. Principal Component Analysis.075. US Patent 6. Mestha and O. Wiley-Interscience. New York. 13. Mikkilineni. and method of generating a reference database. 14. 301–305. Application of principal component analysis and Gaussian mixture models to printer identiﬁcation. 5. S. 5. Color Research & Application. and G. 5. US Patent 5. Bala. 5.D.L. Peterson and M. Hung. 2004. US Patent 7.551.303.000.K.A. Ioka. Color printer characterization adjustment for different substrates.150.556. Mestha. J.272. 28. Loudermilk.333. K. Journal of the Optical Society of America A.R. A. Ishii.584.456. in Proceedings of International Congress of Imaging Science (ICIS’06). L. 725–730. NY.H.525. 18. 2001. and E. 6. 7.838. 6.W.A. Dianat. Mestha. Ramirez.425. Noble and J. Applied Linear Algebra. Mestha and S.K. Hardeberg and F. T.129. AZ. 21. 5. ISBN 0-13-041260-0. New York.T. 9. 1989. G. 28(6). Color printer characterization using a computational geometry approach. 318–322. Mihalyov. 5. Chapter 6. and E. Shaw. Scottsdale. Color measurement instrument. Journal of the Optical Society of America A.853. Jul. Other US Patent 7. Englewood Cliffs. K. Colorimetric calibration in electronic imaging devices using a look-up table model and interpolations. and D. Wada. Salt Lake City. 2008. Liang.451. Feb. L.749.262. 1993.562. 2004. 31. S. R. 1968. 6. Electronic Imaging. 4. Other US Patents in Color Sensor 7.603. G. J.P. 22. F. Y. 6. Daniel. F. Low cost LED based spectrophotometer. Vincent.A.761. US Patent 6. Schweid and L.837.020. Systems and methods for determining spectra using dynamic Karhunen–Loeve algorithms with measurements from LED color sensor. L.T. 31–Nov.037.612. R. 53–61. 17. Mar. D. Skinner. Rochester. 5. 15. Prentice-Hall. On-line model prediction and calibration system for a dynamically varying color reproduction device. Jun. 6. G. E. S. 8. Oct. US Patent 5.671. K.W. Mestha. Dec.059. Springer. and Printer identiﬁcation based on texture features. Allebach. Chiang. Karhunen.174. Oct. Oja. Wang.E.M. 1997.F.A.698.A. Hyvarinen.818. 3. B. Jaaskelainen. Independent Component Analysis. T.P.300. Hubble II.853. 454–467.157.Z. 6.347. Parkkinen.680. T. L. NJ. and A. 2006. May 5. May 7–12. Skinner. 6. pp.W.S. 24.364. US Patent 7. US Patent 7.Three-Dimensional Control of Color Management Systems 423 5. M. . 2007.R.518. Coordinitization of tone reproduction curve in terms of basis functions.377.S.N. Polo. 1993.137. Attorney Docket 20070653.917. 1997. Apparatus and method for measuring color.963.K. 5. 16.567.K.397.541. ISBN 0-387-95442-2. Colorimeter and calibration system.Y. 7. ISBN 0-474-40540-X. Toyooka. Delp. 2007. Mestha. 12. Jun.

E. and R. Viggiano. Oct. The color gamut of halftone reproduction. Proceedings of the American Control Conference. Rotea. 2. Viassolo. Neugebauer. Woolfe.768. pp. 156–166. Journal of Electronic Imaging. in TAGA Proceedings. in 11th Color Imaging Conference: Color Science and Engineering Systems. Robust estimation algorithm for spectral Neugebauer models. 351–385. Apr. Characterization of color printers using robust parameter estimation. Hersch. The use of spectral regression in modeling halftone color printers. 17. Technologies.J. Xia. 4. in Proceedings of IS&T=SID Color Imaging Conference: Transforms and Portability of Color.S. pp. R.S. 73–89. Spaulding. May 1999. Balasubramanian. Balasubramanian. Balasubramanian. 44–62. 29. 27. Vol. pp. pp.D. pp. and A. Jul.–Nov. Salt Lake City. 37. 35. AZ. 2007. A robust estimation algorithm for printer modeling. 543. 36. Dec. 31. A critical review of spectral models applied to binary color printing. Sharma.E. and G. Berns. 2004. 700–716. S. NY. A ﬂexibility study of spectral color reproduction. 1999. Modeling the color of multicolored halftones. 372–375. and S. 1992. 2636–2641. US Patent 7. Wyble and R. IST=OSA Annual Conference. End-to-end color printer calibration by total least squares regression. Nov.J. 40. pp. 39. A uniﬁed model for color prediction of halftoned print. 38. Viggiano. AZ. Optics Imaging in the Information Age. 46(3). M. Proceedings of the IEEE Conference on Decision and Control. J. 2. Jing.S.M. Journal of Imaging Science and Technology. Oct. Nov. 2005. Colorimetric modeling of binary color printers.D. 41. Nov. Berns. pp. 2413. Yang. Journal of Imaging Science and Technology. Vol. Balasubramanian. Proceedings of the SPIE: Color Hard Copy and Graphic Arts IV. NY.M. 2004. D. 863–881. in Proceedings of the Fourth IS&T=SID Color Imaging Conference.A. Gustavson. Imai. 47. . 2003. 43. in NIP20: Proceedings of the IS&T’s International Conference on Digital Printing Technologies. 224–231. Rochester. 2003. Nov. C. Emmel and R. Tekalp. Scottsdale. Rochester. DC. 3. 2003. 42.W. 1990. G. Balasubramanian. 4109–4114.424 Control of Color Imaging Systems: Analysis and Design 25. Braun. 2004. D. 44. G. 28. 4–19. pp. pp. 44(4). Roberts. Oct. K. 1995.H. 2002. Salt Lake City. Hersch. Lana. K. 32–37. 1993. R. Accuracy of various types of Neugebauer models. R. 25(1). Rochester. NY. Bandyopadhyay and T. 3. 647–661. NIP20: Proceedings of the IS&T’s International Conference on Digital Printing Technologies. R. Lena. 26. UT. Vol. D.245.848. Zeitschrift fur wissenschaftliche Photographie Photophysik and Photochemie. Saber. H. reprinted in [22. US patent 6. Modeling ink spreading for color prediction. Spectral model of halftone on a ﬂuorescent substrate. J. 1985.A.E. L. M.R. 7(5). 1996. 2. pp. 327–330. C. 34. Feb. Rotea and C. Lana. Persiev.395. Rotea. The color of halftone tints. A new model of printer characterization. P. Viassolo. Jul. 32. R. 36.–Nov.M. Paul. Color ink model processes for printers. Calibrating a digital printer using a cost function. A printer model for dot-on-dot halftone screens. 405–409. F. M. 1937. Applications. 1995.J. Heuberger. 1996. 425–428. Washington. 33. 19–22. in Proceedings of the TAGA=ISCC. Optimization of the spectral Neugebauer model for printer characterization. and D. in Proceedings of the IEEE International Conference on Image Processing. Z. Color transformations and lookup tables. Journal of Imaging Science and Technology. and D.S. S. M. Couwenhoven. 356–364. 30. 2000. Color Research & Application. 8(2).R. Die theoretischen Grundlagen des Mehrfarbenbuchdrucks. in TAGA Proceedings. UT. Emmel and R. 194–202]. Scottsdale. IEEE Transactions on Image Processing. P. 2000. 237–246. Wyble. P. Rolleston and R. Feb.

Archer. Velde and P. Iino and R. Y. Proceedings of SPIE. Applied Optics.S. in Proceedings of the 7th Annual Technical Meeting TAGA. 267–277. I. 1997.633. TAGA Proceedings. 1998. W. Sep. Methods of computing ink amounts to produce a scale of neutrals for photomechanical reproduction. Pobboravsky. 65. Oct. TAGA Proceedings. 1978.C. Berns. Innovative method for spectral-based printer characterization. May 2005. 25(4). TAGA Proceedings. 40. 26–27. Huang. 26. F.R.265. Rozzi. Method and apparatus for calculating color gamuts. 023008. K. in TAGA Proceedings. 231–234. 51. 42(1).C. 154–172. 233. Prepress system for offset printing. Vol. Building color management modules using linear optimization I. An inverse solution to the Neugebauer equation for the calculation of grey balance requirements. Iino and R.D. Hersch. K. Arney and C. Yule and W. Jan. Mahy and P. May 1955. 63. and P. 7–18.G. 1966.870. Nov. 65–76. US Patent 6. Neugebauer and Demichel: Dependence and independence in n-screen superpositions for colour printing. Arney. 59. 1996. Yule. F. 600.A. 2000. and G. Clapper and J. J. Berns.Three-Dimensional Control of Color Management Systems 425 45.408. 53. 14. 1. Hauser. 1924. Schettini.832. 1977. I. Color separation method. pp. I. IARIGAI 16. TAGA Proceedings. 17(21). 1989. M. 49. pp. Color Research & Application. 52. Ehbets. Delabastita.R. Journal of Imaging Science and Technology. Journal of the Imaging Science and Technology. Zufﬁ. 1184. Pobboravsky and M. Schettini. 208. Yule–Nielsen effect in printing: A physical analysis. 42(1). 64. 43(7). Farbrezeptberechnung. pp. 65–77. E. Modeling the Yule–Nielson effect. Journal of the Optical Society of America. Computation of dot areas required to match a colorimetrically speciﬁed color using the modiﬁed Neugebauer equations.A. Inversion of the Neugebauer equations. K. Rohodes. 14. The penetration of light into paper and its effect on halftone reproduction. H. M. A modiﬁed factorization of the Neugebauer equations to compute accurate mask characteristics for neutral balance. in Procede. Delabastita. Chappuis. 3. 48. H. Reproduction of color with halftone images. 67. Journal of Electronic Imaging. 61. in NIP20: Proceedings of the IS&T’s International . Proceedings of SPIE. Efﬁcient characterization of printing systems for the packaging industry. Hua and K. J. 3376–3383.C. Journal of Imaging Science and Technology. 56. Zufﬁ and R. Farbe þ Lack. 58. pp. US Patent 5. Liu. S. 57. Porc. C. Yule. Spectral modeling of photographic printing based on dye concentration. 60. Amidror and R. W. 2003. 17–21. Mauri. 66.D. desktop color system. P. 79–94. 1991.A. Fifty years of the Neugebauer equations. 365–374. 54. 4663.R. pp. Spectral reﬂectance modiﬁcation of Neugebauer equations.A. 46. 62. in Proceedings of IS&T=SID’s 5th Color Imaging Conference. 21(6).V. The effect of multiple internal reﬂections on the densities of half-tone prints on paper.S. 1996. May 1951. 10–34. D Hasler. 2007. 1981. US Patent 7. Nov. Nielsen.S.109. Zimmermann. 1998. 1997. 1–14. 84(4). Vol. Mahy. pp. 1972. S. Spectral-based printer modeling and characterization. 79–94. 1977. 55. Clapper and J. 1953. Scottsdale. Advanced cellular YNSN printer model. F. Building color management modules using linear optimization II.B. Pearson. Von Schmelzer. Ruckdeschel and O. B. Color Research & Application. R.H. AZ. 47. pp. pp. 50. Demichel. 2002.J. 256–280.

1947. 1986. 71. (Kita 0 529) and US Patent 5. Maltz. Plouffe. R. Y. WA. 448–457. J. New York. S. in IEEE Region 10’s Ninth Annual International Conference. WA. Journal of Imaging Science & Technology. 2003. 27–32. in Proceedings of SPIE. Springﬁeld. US Patent 7. Smooth gray component replacement strategy that utilizes the full device gamut. US Patent 5. SPIE Press. H. FL. 77. A new look at statistical model identiﬁcation. Tsukada and J. Mahy.195.W. S. 410–419. AC-19. 4(3). 85. Akaike. 72. P.579 to Kita et al. UT. 84. Color Technology for Electronic Imaging. IS&T. pp. Proceedings of SPIE.696. 365–374. C. 86.425. 390–393.827. 31–Nov. 80. 12. Kang.R. 1999. H. Sayanagi. M. Springﬁeld. 2001.426 Control of Color Imaging Systems: Analysis and Design Conference on Digital Printing Technologies. Gray component replacement by the Neugebauer equations. 2413.502. G. 78. 3(4). Bala. The black printer. W. 1999. 123–133. Birkenshaw. Image transformation into device dependent color printer description using 4th-order polynomial regression and object oriented programming development of image processing modules. pp. 716–723. in Color Imaging Conference: Transforms and Transportability of Color. 1994.134. Nakamura and K. Tominaga. H.C. 1995. 6–9. Vol. 50–63. 68. 5.. Performing color space conversions with three-dimensional linear interpolation. 357–363. pp. Salt Lake City. 81. in Digital Color Imaging Handbook. Los Alamitos. 1995. 4.R.L. Iga.878. P. 2007. pp. Vol. Proceedings of SPIE. Design of UCR and GCR strategies to reduce moire in color printing. 2414. Color separation method for same. Oct. 82. pp. IS&T. 2.I. Tritton. 75. Billingham. Kang.199. M.R. 1995. M. Y. H. Sharma Ed. RGB-YMCK color conversion by application of the neural networks. 2658. 74. Bellingham.917. Balasubramanian and R. 70. New algorithm for UCR using direct color mapping. (Kita 0 290). Eschbach.411. Printer-related color processing techniques. IS&TPICS Conference. SPIE Press. 1996. 341–352. 5. Tajima. New contributions to the optics of intensely light-scattering materials: Part I. J. and T. ISBN 0-8194-2108-1. 73.508. Balasubramanian and R. Kasson. US Patent 5. R. 88. Nakano. 76.M. 1996. Marcu. Nin. IEEE Transactions on Automatic Control. 1993. in TAGA Proceedings. in IS&T=SID Color Imaging Conference: Transforms & Transportability of Color. in Proceedings of SPIE. Iwata. VA. M. CRC Press. Proceedings of SPIE. 1994. R. K. and S. . 415–424. 79. Hung. 3. pp. VA. Device characterization. Applications of color mixing models to electronic printing.482. 2658.290 to Kita et al. 38. Eschbach. 276–287. 41–44. pp. and J. Scott-Taggart. 253–260. Reducing multiseparation color moire via a variable undercolor removal and gray-component replacement strategy. Proceedings of SPIE.T. pp. Mar. 5. March=April. 1996. A ﬂexible GCR based on CIE L*a*b*. Boca Raton. 1989. Hafner. 1995.636. H. Smooth colorimetric calibration technique utilizing the entire color gamut of CMYK printers. Kita. SPIE Optical Engineering Press. Abe and G. 45(2). 83. Marcu and K. 1994. 1184. Kubelka. 1993. CA. 226–250. and K. IEEE Press. 2004. Murai. 69. Journal of Electronic Imaging. G. Kang. Journal of Electronic Imaging. A method of transformation from CIE L*a*b* to CMY value by a three-layered neural network. 87. 2413. M. A neural network approach for RGB to YMCK color conversion. Chapter 5. Mongeon. 5. Aug. S. 152–160.553. 1973. Arai. Ogatsu. Color control using neural networks and its application. Journal of Electronic Imaging. Journal of the Optical Society of America. 403–429.

Nuyan. 98. Jr. 1999.K. J. 1992. Jan. Gil and L. Subbarayan. and L. and B.S. and G. Reﬁnement of printer transformation using weighted regression. and S. Shepard. P. Gregory. School of Mathematical Sciences. Ogata. NJ. Proceedings of SPIE. US Patent 7.194. 1997. Chong and S. MD.E. 2007. 1996. 95.K. Proceedings of ACM National Conference. L. US Patent Application 20080043264. Nonlinear systems. 99.2.A. pp.M. Dianat.E. 103.K. Khalil. Inc. 517–524..S. Bedford Park. Sep. Salt Lake City. Bennett.2. Webster. Prentice Hall.A.Three-Dimensional Control of Color Management Systems 427 89. 8. Journal of Optical Engineering.F. 42(3). K.H.581.. 5042. 105. Gil. 110. Annaswamy. Viassolo.G. 31. UT. D. 2003. 96. Littlewood and G. Stochastic optimization: Stochastic approximation and simulated annealing. 2000. Gray balance control loop for digital color printing systems. Oct.857. R. Mestha. M. Wang. A two-dimensional interpolation function for irregularly-spaced data. Blended look-up table for printing images with both pictorial and graphical elements.=Dec. Lewis.. Mar. Ed. US Patent 6.397.A.N. 90. 2008 (Extensions to multidimensional smoothing described in Sections 6.C. Method and apparatus for controlling the spectral reﬂectance of a material. NIP20: Proceedings of the IS&T’s International Conference on Digital Printing Technologies. D.A. Controlling the gray component with Pareto-optimal color space transformations.J.E. NIP20: Proceedings of the IS&T’s International Conference on Digital Printing Technologies. Multicolorant process control. S. 101. 108. Flinders U. Robust eigenstructure assignment in state feedback control. ACM Transactions on Graphics. 97. Spall. Mestha. 2658.K. Control System Toolbox for use with MATLAB. NY. Wiley. 625–631. W. D. 499–504. Prentice Hall. 93. Zak.E. J. R. US Patent 5.. R. Aug. L.6. 2002. E. Addressing the problems of color science and management in toner-based digital print. Apr.2 and 6.734. Dianat. Y. Stable Adaptive Systems.K. 3rd edn. Lan. John Wiley & Sons Inc. 1999. Maltz. 2002. ISBN 0-47181240-4. Subbarayan.K. S.. Nov. Australia. 5. . Jul.063. 46(6). 1995. and S. Narendra and A. L. 91. Mestha. Attorney Docket 20061087. Poe and H. 2002.S. 102. US Patent Application 20020105659. Mar. Dianat. 92. A.K. A practical algorithm for the inversion of an experimental input–output color map for color correction. Upper Saddle River. 2008. 20. An Introduction to Optimization. MA. Viturro. Harrington.6. Dover Publications. Drakopoulos. Zhang. Discrete-Time Control Systems. 21(2). Mestha. D. Ma. John Wiley & Sons. A method for classifying a printer gamut into subgamuts for improved spot color accuracy. 8. Journal of Imaging Science and Technology.802. Inc. Optimal Control. New York. F. 5. 2004. and M. pp. 132–174. 94.A. The Math Works Inc.3). 2005. 31–Nov. W. 109. Numerical Analysis Report NA=2=83. Feb. Spot color controls and method. pp. S. Rozzi. 2002. May 2005. US Patent 5. 104. 100. Vol. Littlewood. ISBN 0-13-067389-7. ISBN-10-0-486-442268. A.052.K. Dec. Pareto-optimal formulations for cost versus colorimetric accuracy trade-offs in printer color management. pp. 334–340.. 107. Balasubramanian and M. Baltimore.. 106.R. ISBN 9-13-034281-5. 1996. K. US Patent Application. S. A. Mestha and S. Hoffmann. TRC smoothing algorithm to improve image contours in 1-D color controls. 1967. J. in Encyclopedia of Electrical and Electronics Engineering. New York. New York. 1986..L. Maltz. Mineola. H. 529–542. 342–346. Kautzky and N. Nichols. 18. Gamut-preserving color imaging.

M. 2007. 15(1). 128. 45(3). Jan. and J. 144–148.R. Montag and M. Braun. Montag and M. Color Gamut Mapping. 7(4). Gamut mapping squeezing the most out of your color system. Spaulding. E. John Wiley & Sons.R. Allebach. Simon. IEEE Transactions on Image Processing. J. R. P.306. . Eschbach. Journal of Imaging Technology. E.W. 119. Oct. The fundamentals of gamut mapping: A survey. 118. T. P. Wolski. Balasubramanian. Walowit. Morovic and M. Lou. M. Oct. 16(5).J. Woolfe. 23. 1774–1783. 1999. 2005. J.P. 1998. Ronnier Luo. Elison. 20(5). Stone and W. Nov. Journal of Electronic Imaging. Implementation of real time color gamut mapping using neural network. 112. E. 1995. 249–292. K. Stone. Spaulding. 11. Calculating medium and image gamut boundaries for gamut mapping. J. and C. Color gamut mapping for accurately mapping certain critical colors and corresponding transforming of nearby colors and enhancing global smoothness. S. Scottsdale. Feb.M. Granger.030. June 28–30. Color gamut mapping and the printing of digital color images. Berns. R. 116. Finland.S. CA. Sep. 1997. Han. Lee and D. 57–61.D.J. Proceedings of SPIE. Gentile. Continuity of gamut mapping algorithms. W. and G. Allebach.S.E. Zolliker and K.S.M. pp.R. 13004. Journal of Imaging Science and Technology.C. 394–401.C. Psychophysical evaluation of gamut mapping techniques using simple rendered images and artiﬁcial gamut boundaries. Hordley. New York. Wang. 6. G. 2414. and J. Mar. AZ. D. 2000. 113. Morovic. 123. 122. 283–290. Color Research & Application. IEEE Transactions on Image Processing. G. Sullivan. 2001. K. ISBN-10: 0470030321. pp. 114.E. and Applications.D. Hoshino and R. 1988. Marshall. 1988. 2006.-S. Fairchild. The fundamentals of Gamut mapping. 121.B. ACM Transactions on Graphics. 1995. San Jose. R. 1995. Color gamut mapping and the printing of digital color images. 2003. Oct. H. IS&T=SID 7th Color Imaging Conference. Color Research & Application. Systems.-C. S. Bouman. 124. AZ.C. 2005 IEEE Mid-Summer Workshop on Soft Computing in Industrial Applications. in IS&T=SPIE Symposium on Electronic Imaging: Science & Technology. M. Espoo. 89–92. Hung and R.J.–Mar. 2000. Scottsdale. and J. Jul. 1909. The Sixth Color Imaging Conference: Color Science. A comparison of techniques for color gamut mismatch compensation. Selectively pleasing color gamut mapping in a color computer graphics system. M. Couwenhoven. 125. 129. 1994. Braun.D. 176–181. US Patent 7. 5–10. 7(4). E.A. US Patent 6. K. Fairchild. 120. 2000. 977–989. 2007. 1997.B.P. Proceedings of SPIE. 1990. UltraColor: A new gamut mapping strategy. Color gamut mapping using a cost function. 115.R. 27. US Patent 5. 1993. UK. 131.D. Systems and Applications. J. J. Finlayson and S. Helsinki University of Technology. and R. Beatty. Luo.611. in IS&T and SID’s 2nd Color Imaging Conference: Color Science.625. ACM Transactions on Graphics. Morovic and M. Development and evaluation of six gamut-mapping algorithms for pictorial images.422. 3. Gamut mapping for hard copy using the ATD color space. Cowan.239. 117. Color gamut mapping techniques for color hard copy images. Harrington. Cowan. 152–165. Morovic and M. Jackson. 130. 25(6). 126. pp. 127. Improving gamut mapping color constancy. Color Research & Application. 9(10). E. and Y. Berns.W. Gamut mapping: Evaluation of chroma clipping techniques for three destination gamuts. Stokes.N.428 Control of Color Imaging Systems: Analysis and Design 111. Determination of constant hue loci for a CRT gamut and their predictions using color appearance spaces. Oct.

2000. 566–569. Mar. 664–672. Jun. 10. D. Gil. R. Aug. 2005. Oct. Zolliker and K. Canada. and I. and M. Jun. 134. 16(3). Elad. Space-dependent color gamut mapping: A variational approach.K. BC. 135. Sobel. Merit based gamut mapping in a color management system. Extraction of image gamut surface and calculation of its volume. Simon. A. Maltz. Retaining local image information in gamut mapping algorithms. IEEE Transactions on Image Processing. 2008. P. Saito and H. Shaked. US Patent Application. NIP16: Proceedings of the IS&T’s International Conference on Digital Printing Technologies. Kotera. 136. Mestha. Kimmel.E. 133. 2007. pp. Attorney Docket 20071654. 14(6). ASTM International. ASTM E308–01. Vancouver. R. L. 796–803. M. . 2001. IEEE Transactions on Image Processing. Standard practice for computing the colors of objects by using the CIE system.Three-Dimensional Control of Color Management Systems 429 132.

.

8 One-Dimensional. At the end of the chapter. yellow. customer logo colors. These functions involve the printing. 431 . a publisher might want to print millions of copies of a book on dozens of print engines. colors in a customer’s proprietary marked patterns. For high-quality consistent color rendition. They are normally executed on an ‘‘as needed’’ basis at a lower frequency than the internal control functions. from job-to-job. magenta. Proﬁling and calibration are two important control functions normally implemented external to the print engine. Figure 8. The proﬁling and calibration of a conventional four-color (cyan. We show both model-based and (more sophisticated) control-based techniques. The 1-D=2-D CMYK maps are used to account for device variability over time. CMYK to CMYK calibration map.1 INTRODUCTION A key attribute of high-quality color printing is color consistency within a job. or customer-deﬁned colors in the form of an index color table. that is. and from printer-to-printer.1 shows a schematic of the image-processing path with the destination proﬁle. Print engine colors drift. we extend the control-based methods to generate an accurate spot-color recipe in device-speciﬁc space (CMYK) to render ﬁxed colors that may be Pantone colors. One risk is that the different print engines may produce copies that appear different. and black) digital printer involves (1) generating a three-dimensional (3-D) proﬁling lookup table (LUT) for mapping the device-independent L*a*b* or XYZ color space to device-dependent CMYK color space based on measurement of color patches. and (2) constructing device tone reproduction curves (TRCs) that are one-dimensional (1-D) or two-dimensional (2-D) input CMYK to output CMYK maps. and a print engine with numerous details omitted for clarity. and processing of control patches for a particular paper and halftone screen. the processes internal and external to the print engines have to be controlled at regular print intervals. measuring. This chapter provides detailed algorithms and methodology to generate 1-D=2-D calibration maps. as described in Chapter 7. and Spot-Color Management and Control Methods 8. Two-Dimensional. both in device-speciﬁc space. For example. in the digital front end (DFE).

sRGB). This includes proﬁling and calibration of different devices to provide nearly identical reproduction across them. The camera is generally calibrated and hence it complies to some calibrated RGB space (e.432 Input image transformed to L*a*b* using source profile Control of Color Imaging Systems: Analysis and Design CMY 3-D LUT UCR/GCR CMYK TRC (gray balance.2 (See color insert following page 428. if an image is scanned by a scanner.2 PRINCIPLES OF COLOR MANAGEMENT Consider a digital imaging system with variety of color devices such as a digital camera. digital printer. The image captured by the digital camera is in RGB color space. Our goal is to make the reproduction process as accurate as possible (within physical limits) so that output is device. etc. image scanner.) Image reproduction system. L*a*b* (UCR: under color removal. cathode ray tube (CRT) or liquid crystal display (LCD) monitor. channel-independent linearization. This is called what you see is what you get (WYSIWYG). Obviously. For example. A color management system will make the color conversion and reproduction among these devices as accurate as possible. For example. where a digital camera captures an image that is displayed on an LCD monitor and then printed on a digital printer. consider the image reproduction system shown in Figure 8. displayed on an LCD monitor. 8. and offset press.1 Image path with input and output in device-independent color space. .. GCR: gray component replacement). These RGB values are generally displayed for visualization on a computer FIGURE 8.g. we would like the printed image to be perceived as being identical to what we see on the monitor. this is not possible across all the imaging systems due largely to physical limitations on reproducibility of each device.) Calibrated CMYK Output image shown in L*a*b* Hardcopy Marker Halftoning Print engine FIGURE 8. media.2. and then printed on a digital printer. and halftone independent.

**One-Dimensional, Two-Dimensional, and Spot-Color
**

Uncalibrated camera RGB Calibrated Camera Camera camera RGB calibration profiling RGB RGB RGB L*a*b*

433

L*a*b* or XYZ (PCS)

Monitor profiling L*a*b* RGB

Monitor calibration RGB RGB

Monitor Monitor RGB

Hard copy Printer

Printer calibration CMYK CMYK

Printer profiling L*a*b* CMYK

Printer CMYK

FIGURE 8.3 Color reproduction system (a simple example) (PCS: proﬁle connection space).

monitor (e.g., an LCD monitor, CRT monitor, etc.). Typically, monitors are also proﬁled and calibrated. To produce a hardcopy, media, such as coated or glossy paper, the image ﬁles with camera RGB values are sent to the printer. As described in Chapter 7, there is a common universal color space (e.g., proﬁle connection space (PCS) in international color consortium (ICC) architecture) that is used to communicate color values between various devices. This space is a device-independent color space (L*a*b* or XYZ). Color capability of each device can be universally understood, reviewed, and compared normally in device-independent color space. By proﬁling and=or calibrating a device with multidimensional or 1-D LUTs, it is possible to convert the device-dependent colors into a device-independent color space. An example of this ﬂow is shown schematically in Figure 8.3.

**8.3 ONE-DIMENSIONAL GRAY-BALANCE CALIBRATION
**

Gray-balance calibration is a process by which 1-D LUTs are produced for the separation colors (e.g., CMY with K ¼ 0) using the measurement of output color near neutral. Grayness is an indication of how clean a process color is in a printing system as compared to a theoretical ideal. A raw print engine is normally not gray balanced. Based on appearance, neutral gray balance is often deﬁned as the

434

Control of Color Imaging Systems: Analysis and Design

**TABLE 8.1 G7 Gray-Balance Chart for a Commercial-Coated Paper
**

CMY (%) C 0 12.5 25.1 37.3 49.8 62.7 75.3 M 0 9 18.8 29 40 52.9 66.3 Y 0 9 18.8 29 40 52.9 66.3 a* 0 0 0 0 0 0 0 Nominal CIELab b* À2 À2 À2 À2 À2 À1.5 À1

color of a 50% black ink tint on paper. To ensure the best chance of matching the output of devices with different colored papers or black ink, GRACoL Standards Organization [1] deﬁnes neutral gray balance in colorimetric terms with non-equal CMY as [50 C, 40 M, 40 Y] ¼ [a* ¼ 0.0, b* ¼ À2] with a tolerance of Æ 0.5 for a* and Æ1.0 for b*. Since most people prefer a slightly bluer gray, b* is set close to À2.0 instead of 0, as in neutral color. A similar deﬁnition is speciﬁed by GRACoL standards, as shown in Table 8.1 for other CMY area coverages on a commercialcoated paper, where paper is near neutral with a* ¼ 0.0 and b* ¼ À2. The deﬁnition of gray balance for different paper types is still under discussion by the Standards Organization. However, in Ref. [1], for gray balance, target a* and b* values are shown for a subset of nonstandard paper. If we disregard the effects of paper, then a good gray has zero chroma (i.e., a* ¼ b* ¼ 0.0). When equal amounts of cyan, magenta, and yellow are printed on white paper, a well-balanced printer should produce an equivalent neutral gray. However, a brownish color rather than a neutral gray often occurs due to paper effects. As in GRACoL standards, the CMY values for one or all three of the colors may need readjustment to properly reproduce the gray scale within standards. Rather than trying to develop a gray-balance approach for various deﬁnitions, we show a general approach in which the gray-balance calibration is performed to make the engine produce equivalent neutral gray with equal amounts of CMY colors [2,3]. For example, the equivalent neutral gray means that input digital count of C ¼ M ¼ Y ¼ 40 should produce gray with L* ¼ 100 À 100 Â 40 ¼ 84:3 and 255 a* ¼ b* ¼ 0. The K separation does not come into this deﬁnition. To have a graybalanced device, the input CMY digital counts have to be passed through a transformation LUTs to produce device C0 M0 Y0 , which is then used to make prints. The three 1-D transformations C0 ¼ fc(C), M0 ¼ fc(M), and Y0 ¼ fc(Y) that map the input CMY ¼ [C M Y]T to the device CMY ¼ [C0 M0 Y0 ]T for every digital count (or area coverage) are called tone reproduction curves. In practice, these three TRCs are given by three 1-D LUTs. Each LUT has 256 entries for input digital counts of 0–255. This should not to be confused with channel-wise linearization, as in the paper-based level 3 controls described in Chapter 9 and later in this chapter. Such 1-D TRCs

One-Dimensional, Two-Dimensional, and Spot-Color

435

linearize each separation independently, and, therefore, do not always yield a neutral gray balance for C ¼ M ¼ Y. Since the TRCs (gray-balanced or 1-D channel-wise linearized) contain all the possible entries and are independent of each other, each separation can be processed on a real-time basis without interpolation after RIPping. If a color sensing instrument becomes available, then the construction=generation of the gray-balance TRCs and processing of each pixel with the updated TRCs can also be used as a means to improve the stability of neutral colors. Whereas, when real-time processing is applied with 1-D channel-wise TRCs, we can only stabilize colors along the separations (not neutral colors, such as C ¼ M ¼ Y). Hence, a system can be engineered to automatically gray balance the print engine and to automatically linearize the separations independently to overcome environmental and other disturbances that cause color variations on paper. When 1-D channel-wise linearizations are used in conjunction with a gray-balance calibration, proper separation rules have to be applied while creating and using the TRCs to avoid undesirable loop interactions.

**8.4 TWO-DIMENSIONAL CALIBRATION
**

With 1-D gray-balanced TRC LUTs, as described in Section 8.3, we can achieve color balance to neutral colors by constantly monitoring the colors along the neutral axis (i.e., the L* axis in the L*a*b* color space) and producing color-balanced TRCs. This will provide good control along a sensitive color axis. Although, this approach can make other colors reasonably accurate, it does not control colors everywhere in the color space like the 3-D color control approaches described in Chapter 7. Accordingly, one might ask the following question: Why not just do 3-D color control more frequently to stabilize the color output of engines instead of simply updating gray-balance TRCs on a more frequent basis? There are three reasons why this strategy is not preferred: (1) processing of 3-D LUTs for high-speed RIPping is not cost effective when compared to processing 1-D graybalance TRCs, (2) memory requirements are very minor for 8-bit TRC processing (256 bytes of memory per separation TRC), and (3) the number of levels that can be corrected in a 3-D LUT is limited along each separation axis to a much smaller number (e.g., 33 in a 333 proﬁle LUT) as compared to 256 levels in the 1-D TRCs, thus limiting the quality levels for rendering neutral colors. A full resolution 3-D LUT with full lookup can avoid the shortfall in levels, but can be prohibitively large (a full 3-D LUT size for a 8-bit system would be 3 Â (256)3 bytes ¼ 49.75 M Bytes of storage), especially when customized for each media and halftone screen. Gray balance alone can introduce undesirable hue shifts in the reproductions of sweeps from white to secondary colors, such as those colors lying along the red axis (M ¼ Y and C ¼ K ¼ 0) as well as those along the blue and green axes. On the other hand, with a 2-D calibration [4], we can achieve gray balance as well as control over colors lying on other critical axes in the color space. In Section 8.6.3, the 2-D calibration approach is discussed. In this case, the TRCs are 2-D and the LUTs representing each TRC would have 2562 ¼ 65,536 entries, quite reasonable as compared to full resolution 3-D LUTs. Obviously, as we try to control more axes in the color space, the LUT size increases. The 2-D TRCs offer signiﬁcant levels to

436

Control of Color Imaging Systems: Analysis and Design

control as compared to 1-D channel-wise or 1-D gray-balance TRCs and hence can provide more accurate color rendition than 1-D TRCs, but are still less accurate than 3-D LUTs [5].

**8.5 ONE-DIMENSIONAL AND TWO-DIMENSIONAL PRINTER CALIBRATION USING PRINTER MODELS
**

In this section, we show how to construct the 1-D and 2-D printer calibration TRC LUTs if an updated printer model is given. The model can be in any of the forms described in Chapters 7 or 10 or even a simple input–output LUT obtained experimentally while characterizing the printer. We ﬁrst discuss the algorithms for 1-D calibration, which involves both 1-D channel-wise linearization and 1-D gray balance to equivalent neutral targets.

**8.5.1 ONE-DIMENSIONAL CHANNEL-WISE (INDEPENDENT) CALIBRATION
**

In channel-independent calibration, each channel is independently linearized. By * linearization, we would like to make the DEab obtained by measuring color patches from paper to be linear so that the channel-linearized printer can emulate an ideal * * printer, which has the characteristic of linearized DEab from paper. The DEab from paper is the Euclidean norm between target color and the paper white in the deviceindependent color space (L*a*b*). The DE2000 metric is not normally used for the paper-based 1-D channel-wise calibration. As an example, let us assume that we choose the cyan channel for linearization and printed N cyan patches (M ¼ Y ¼ K ¼ 0) that represent the step wedge with the cyan digital count of d0 ¼ 0, d1, d2, . . . , dNÀ1 ¼ 255, and measured the L*a*b* values of each cyan patch using a color sensor. Let the corresponding measured L*a*b* value of the ith patch be (L*a*b*)i. Note the ﬁrst patch is the paper white which * means that (L*a*b*)0 is the L*a*b* of the paper. We now compute the DEab from paper for each patch as follows: DEi ¼ k(Lab)i À (Lab)0k i ¼ 0, 1, . . . , N À 1 (8:1)

The function f(.) is formed by normalizing the DE values to obtain a value of 255 for the cyan value of 255. It is deﬁned as di ! f ðdi Þ ¼ 255DEi maxi ðDEi Þ (8:2)

To linearize the printer, we need DEi to be a linear function of di. To achieve this, we ﬁnd the inverse of the transformation given by Equation 8.2, which is the TRC of the cyan channel. That TRC can be represented by following equation: Cyan TRC: c ¼ f À1 (C) (8:3)

One-Dimensional, Two-Dimensional, and Spot-Color

437

We repeat the above process for all the other three channels to obtain the TRCs for the remaining magenta, yellow, and black channels by printing and measuring corresponding color patches. Instead of step wedges, pure sweeps covering 0–255 digital counts can also be used for channel-wise linearization. In such cases, the measurement aperture of the sensing process becomes more important than the gradient of the sweep. The following example shows the 1-D channel-independent linearization process for the cyan channel using patches representing step wedges. Example 8.1

Assume that N ¼ 11 Cyan patches with digital counts of 0, 25, 50, 75, 100, 125, 150, 175, 200, 225, and 255 are printed on a color digital printer. The measured L*a*b* values of these 11 patches and their corresponding DE from paper are shown in Table 8.2. Note that the ﬁrst row, which corresponds to CMYK ¼ [0 0 0 0], is the paper white. Therefore, all the DE values are measured with respect to this reference. In practice, the L*a*b* of different spots on the paper can be measured and the average of all these L*a*b* values used as the reference paper white to avoid errors in the measurement process. The function f(d) is obtained by dividing the DE values by its maximum and multiplying the results by 255. The function is shown in the last column of Table 8.2. The TRC is obtained by ﬁnding the inverse of the function f(d). The inverse can be simply obtained by interchanging the dependent and independent variables as shown in Table 8.3. Also see Section 9.9 for additional information about ﬁnding an inverse for different conditions. The TRC for the digital counts between the data points of Table 8.3 is obtained by curve ﬁtting or interpolation. The function f(d) and the resulting TRC for the cyan channel are shown in Figure 8.4. In this ﬁgure, the nodes represent the measured points and the solid curve is obtained with linear interpolation. A higher-order interpolation approach can be adopted based on the need.

**TABLE 8.2 Measured L*a*b* Values of Cyan Patches
**

d 0 25 50 75 100 125 150 175 200 225 255 L* 97.81 94.24 90.13 86.67 82.23 77.87 74.44 70.45 67.34 63.43 57.50 a* 0.004 À6.43 À9.52 À13.08 À16.71 À18.87 À20.29 À24.03 À27.49 À31.05 À35.59 b* À0.22 À9.78 À14.96 À16.07 À19.90 À21.34 À25.32 À34.33 À39.25 À44.05 À49.97 * DEab 0 12.06 19.15 23.37 30.15 34.64 39.85 49.89 56.63 63.77 73.26 f(d) 0 42 66.68 81.37 104.97 120.57 138.71 173.68 197.14 221.99 255

438

Control of Color Imaging Systems: Analysis and Design

**TABLE 8.3 Cyan TRC Function
**

d 0 42 66.68 81.37 104.97 120.57 138.71 173.68 197.14 221.99 255 TRC ¼ f À1(d) 0 25 50 75 100 125 150 175 200 225 255

250 f (d ) TRC 45° Line

200

150 TRC 100 50 0 0

50

100 150 Digital input C

200

250

FIGURE 8.4

Channel-independent cyan TRC.

**8.5.2 GRAY-BALANCED CALIBRATION
**

Recall that gray-balance calibration is performed (i.e., gray-balance TRCs for cyan, magenta, and yellow channels are generated) to make the print engine produce equivalent neutral gray accurately with equal amounts of CMY colors. That is, after gray-balance calibration, an input with a digital count of C¼M¼Y ¼d and K¼0 (8:4)

One-Dimensional, Two-Dimensional, and Spot-Color

439

produces gray level with L* ¼ 100 À d Â 100, 255 a* ¼ b* ¼ 0 (8:5)

The TRC for the black channel is also constructed as a second step to perform channel-wise linearization. The main reason for the adaptation of the gray-balance calibration for the CMY channels is the fact that the human visual system is very sensitive to colors near the neutral axis. Small color differences near the neutral are more noticeable than larger color differences far from it. Once the gray balance is achieved with TRCs and the black channel is linearized, the printer will tend to make other colors containing combinations of CMYK lot more pleasing and closer to what is required by the press (see the images in Figure 8.5 with and without gray balance). However, the choice of gray-balance reference or target values matters a lot with respect to the neutral appearance [1]. If the gray-balance reference is altered, then the color balance will also be altered and may result in an error in color. In system theoretic terms, the outcome of the gray-balance calibration can be thought of as a way of generating inverse maps to linearize the printer along the chosen reference=target axis. For example, when the target values are selected on the neutral axis, the printer with a gray-balanced inverse map will be linear to

FIGURE 8.5 (See color insert following page 428.) gray-balanced print engine.

CMYK test images with and without

440

Control of Color Imaging Systems: Analysis and Design

neutral colors, but is not linear for other regions of the color axes in the 3-D CMY color space. If the reference values are off-neutral, then colors are adjusted by changing CMY to follow the off-neutral axis, effectively making the printer linear to the off-neutral reference axis. To perform gray-balance calibration, we take the following steps: (1) Generate a set of CMY color patches in and around the neutral axis (our reference axis). Print and measure the corresponding L*a*b* values and form a printer forward map. This can be achieved either by creating M3 uniformly spaced CMY patches to form a LUT or by measuring fewer patches to ﬁt a model and then use the model to create the desired LUT (i.e., uniformly spaced CMY ! L*a*b* forward map). (2) Utilize interpolation (ICI or tetrahedral) or a regression method to compute N CMY values that yield a neutral measurement (a* ¼ b* ¼ 0) for the L* * values spanning the dynamic range of the printer (L1 , L*, . . . , L* ). 2 N If the CMY gamut coverage is such that the a* ¼ b* ¼ 0 axis is not within the CMY gamut, then create smooth off-neutral reference values. In particular, this may be required for dark regions of the color space. Let the CMY values corresponding to Li* be CMYi ¼ [Ci Mi Yi]. Now compute 255 100 À Li* Di* ¼ 100 for i ¼ 1, 2, . . . , N (8:6)

(3) Create the three TRCs using the following functions: (a) Cyan TRC: Cout ¼ f1 ðCin Þ, where f1 : Di* ! Ci (b) Magenta TRC: Mout ¼ f2 ðMin Þ, (c) Yellow TRC: Yout ¼ f3 ðYin Þ, where f3 : Di* ! Yi (8:9) where f2 : D* ! Mi i (8:8) (8:7)

Note that the above three TRCs are only deﬁned for N input digital counts. Linear or cubic interpolation is used to obtain all the values from 0 to 255, in steps of one digital count. Also, TRCs should be smooth and monotonically increasing functions of their arguments. If required, 1-D smoothing techniques described in Section 6.6.2.1 are used to generate smooth TRCs for all three channels.

One-Dimensional, Two-Dimensional, and Spot-Color

441

Example 8.2

Assume that a 9 Â 9 Â 9 CMY to L*a*b* characterization LUT is measured for a digital color printer. This LUT is used to compute the CMY values corresponding to the gray values of ai* ¼ bi* ¼ 0, Li* ¼ 15 þ 10i, i ¼ 1, 2, . . . ,6. The ICI algorithm (Section 6.4.3) is used to ﬁnd the right combination of the CMY values. The results are shown in Table 8.4. Note that since ICI is an iterative technique, the iterations are terminated once the error or number of iterations reach a predeﬁned threshold. Hence, exact gray with ai* ¼ bi* ¼ 0 may not be achieved. Table 8.5 shows the three gray-balance TRC functions. Figures 8.6 through 8.8 show the plots of the cyan, magenta, and yellow gray-balanced TRCs, respectively. The available seven data points on the TRCs are marked by a circle and the other values are obtained through linear interpolation (for internal points) or linear extrapolation (for external points). Note that since smooth TRCs are constructed, the ﬁnal TRCs do not pass through the data points, particularly in the shadow region. This has the effect of making gray balance less accurate in those regions, but may be essential to reproduce smooth, contourless shadow colors.

**TABLE 8.4 Input CMY Values that Generate Gray
**

L* 25 35 45 55 65 75 a* 0.005 0 0.003 À0.002 0.002 0.0017 b* 0.005 À0.002 0.004 À0.003 0.003 0.001 C 235.25 203.44 164.89 131.58 97.88 68.48 M 237.18 200.61 169.45 129.78 104.02 70.85 Y 221.58 187.41 155.1 132.97 104.85 68.73

**TABLE 8.5 Three Gray-Balance TRCs
**

D* ¼ Cin, Min, Yin 191 166 140 115 89 64 Cout 235.25 203.44 164.89 131.58 97.88 68.48 Mout 237.18 200.61 169.45 129.78 104.02 70.85 Yout 221.58 187.41 155.1 132.97 104.85 68.73

442

250

Control of Color Imaging Systems: Analysis and Design

200 Cyan TRC 150 Cout 100 50 0 0

50

100 Cin

150

200

250

FIGURE 8.6 Cyan gray-balanced TRC.

250

200

Magenta TRC

150 Mout 100 50 0 0

50

100

Min

150

200

250

FIGURE 8.7

Magenta gray-balanced TRC.

**8.5.3 TWO-DIMENSIONAL CALIBRATION
**

One-dimensional calibration is limited to color balancing the printer to one of the color * axes or to achieve a linearized DEab response with respect to paper. Achieving both simultaneously with TRCs is not simple. A full 3-D function with input Cin, Min, and Yin to output Cout, Mout, and Yout is being applied in many color management systems as an alternative to 1-D calibration. In this case, the three 3-D TRCs (or LUTs) would be

**One-Dimensional, Two-Dimensional, and Spot-Color
**

250

443

200 Yellow TRC 150 Yout 100 50 0 0

50

100 Yin

150

200

250

FIGURE 8.8 Yellow gray-balanced TRC.

Cout ¼ f1 (Cin , Min , Yin ) Mout ¼ f2 (Cin , Min , Yin ) Yout ¼ f3 (Cin , Min , Yin )

(8:10) (8:11) (8:12)

where Cin, Min, and Yin are the inputs of the 3-D map and Cout, Mout, and Yout are the outputs of the 3-D map, which are the actual device CMY values sent to the printer. Alternatively, in 2-D calibration, three 2-D TRCs are constructed as follows: Cout ¼ f1 (Cin , Min þ Yin ) Mout ¼ f2 (Min , Cin þ Yin ) Yout ¼ f3 (Yin , Cin þ Min ) (8:13) (8:14) (8:15)

For example, the cyan TRC given by Equation 8.13 suggests that the calibrated cyan value is a function of input cyan and sum of input magenta and yellow. The domain of this function or LUT is illustrated in Figure 8.9, and the implementation of Equations 8.13 through 8.15 is represented by the block diagram of Figure 8.10. The advantage of 2-D calibration over 1-D is that we can obtain good control over ﬁve color axes, as shown in Figure 8.9. By controlling the gray axis, we can achieve good gray balance; controlling the cyan primary axis, we can achieve channel-independent linearization. By controlling other axes such as primary to black, we can linearize the printer along those axes, achieving better control of the printer gamut. The process of linearization along these axes is similar to the channelindependent linearization.

444

Cin 255

Control of Color Imaging Systems: Analysis and Design

Secondary axis

C = M, Y = 0

C=M=Y Gray axis

Cyan primary axis M=Y=0 0 ≤ C ≤ 255

Secondary to black

M = Y = 255 0 ≤ C ≤ 255 Min + Yin

0

255

510 Primary to black C = Y, M = 255

FIGURE 8.9 Two-dimensional cyan TRC with different control axes.

2-D LUT C + C M+Y

M + Y +

M C+Y Y C+M

P

L*a*b* out

FIGURE 8.10

Two-dimensional calibration.

Example 8.3

Assume that a 13 Â 13 Â 13 CMY to L*a*b* characterization LUT is measured for a digital color printer. The 2-D LUTs for C, M, and Y were ﬁlled in by controlling different axis shown in Figure 8.9. Fifteen grid points are chosen on each control axis. The scatter plot of these grid points for the 2-D cyan TRC in (C, M þ Y) plane is shown in Figure 8.11. The corresponding 2-D LUT for the cyan TRC corresponding to the control grid points of Figure 8.11 is shown in Table 8.6. This is obtained using the ICI algorithm.

**One-Dimensional, Two-Dimensional, and Spot-Color
**

C 250

445

200

150

100

50

0 0 50 100 150 200 250 300 350 400 450 500

M+Y

FIGURE 8.11 Scatter plot of grid control point for 2-D cyan TRC in (C, M þ Y) plane. Note that the LUT includes the control grid points of Figure 8.11. The complete LUT for every point in the (C, M þ Y) plane is constructed using interpolation= smoothing. Similar LUTs are constructed for magenta and yellow TRCs. To evaluate the performance of 2-D calibration against 1-D, a total of 180 in-gamut patches were generated. The patch values are processed through the calibration LUTs and the simulated printer for both 1-D and 2-D calibrations. The accuracy results are compiled in Table 8.7. Two-dimensional calibration outper* forms 1-D with respect to mean and 95th percentile DEab .

**8.6 ONE-DIMENSIONAL AND TWO-DIMENSIONAL PRINTER CALIBRATION WITH STATE-FEEDBACK METHODS
**

The approach of the previous section regarding the gray-balance and channelindependent calibrations is based on printer model constructed using a high-resolution LUT or an accurate printer model. In this section, we show how control approaches using state-feedback methods [2,3] can be used to design 1-D and 2-D calibration TRCs. Other, more robust, control approaches are also available [6]. Consider the gray-balance control loop with the gain-weighted MIMO (multiinput and multi-output) integral controller shown in Figure 8.12. In this ﬁgure, x(k) ¼ [L*(k) a*(k) b*(k)]T are the states, as deﬁned by the state variable model for a CMY ! L*a*b* printer described in Section 7.5.2.2, and, in this case, the states are also the output of the system. The outputs are the measured L*a*b* values from the * * * color sensor. V(k) ¼ [dC(k) dM(k) dY(k)]T. r ¼ [Ltarget atarget btarget ]T is the target L*a*b*, and K is the 3 Â 3 feedback gain matrix. The printer in this loop can be an (accurate) printer model or an actual printer. The sensor can be an inline or ofﬂine spectrophotometer. If an inline sensor is used, then the processes can be automated to

446 Control of Color Imaging Systems: Analysis and Design TABLE 8.6 Two-Dimensional LUT for the 2-D Cyan TRC Cin 0 0 0 0 0 0 0 0 0 0 0 0 0 3 8 12 16 20 24 29 33 37 41 45 50 54 58 62 66 71 75 79 83 87 92 96 100 104 108 113 117 121 125 Min þ Yin 0 42 84 126 168 210 252 294 336 378 420 462 504 510 0 12 32 275 510 0 33 74 296 510 0 54 116 317 510 0 75 158 338 510 0 96 200 359 510 0 117 242 380 Cout 0 16 34 51 70 89 110 130 151 174 198 224 251 251 6 9 20 231 224 23 26 38 206 197 39 43 57 179 169 56 60 75 151 142 72 76 92 124 115 89 93 111 101 Cin 129 134 138 142 146 150 155 159 163 167 171 176 180 184 188 192 197 201 205 209 213 218 222 226 230 234 239 243 247 251 255 255 255 255 255 255 255 255 255 255 255 255 255 Min þ Yin 510 0 138 284 401 510 0 159 326 422 510 0 180 368 443 510 0 201 410 464 510 0 222 452 485 510 0 243 494 506 4 46 88 130 172 214 256 298 340 382 424 466 508 Cout 95 106 114 144 86 79 131 137 181 70 63 157 162 212 54 47 183 188 235 38 31 209 214 248 21 16 235 240 253 3 253 234 215 194 173 151 128 107 85 64 42 21 1 .

12 1. L*.One-Dimensional. these loops will be iterating on the printer. We have shown the above control loop for one color patch. If an actual printer is used. we can select N number of patches using the nominal CMY values obtained for each of the patch gray-balance targets. The nominal CMY values are constructed using any of the following procedure: (1) When a printer model is not available. As a result of this. it can be a semiautomated system.e.7 Round-Trip Accuracy Numbers for 1-D and 2-D Calibrations Calibration Type One-dimensional gray-balance calibration One-dimensional channel-independent calibration Two-dimensional calibration * * Mean DEab 95th Percentile DEab 5. To construct three TRCs. If a printer model is available. we would have N number of control loops running in parallel.5.32 Ltarget r = atarget btarget + _ K u(k) Integrator V(k) + + x(k) Printer Sensor Nominal CMY values FIGURE 8. compute CMY for the L*a*b* graybalance target using the ICI algorithm.06 5.13 3.12 Gray-balance control loop.23 9. Two-Dimensional.2.16 by knowing the gray-balance target for the chosen patch (i. and b* values): (100 À L*)*255 .2. use C ¼ M ¼ Y ¼ d and K ¼ 0 where d is computed from Equation 8.. a*. . 100 a* ¼ b* ¼ gray-balance target (or 0 for equivalent neutral) d¼ (8:16) (2) When a printer model is available. Otherwise. then the control iterations are carried out as described in Section 7. and Spot-Color 447 TABLE 8. eliminate the need for operator intervention.75 8.

[7].17 can be written for print number k as x(k) ¼ BV(k À 1) þ x0 Subtracting Equation 8. Using the state feedback. can then be designed using pole-placement or optimal control techniques. as described in Chapter 6. u(k) ¼ K[r À x(k)].17 results in x(k þ 1) ¼ BV(k) þ x0 ¼ BV(k À 1) þ Bu(k) þ x0 (8:20) (8:19) Assuming that the Jacobian is not changing between two consecutive prints. K. Equation 8.21 from Equation 8. we can approximately represent the dynamics of the single-color reproduction system using a ﬁrst-order ﬁnite difference equation given by x(k þ 1) ¼ BV(k) þ x0 where x0 is the L*a*b* vector corresponding to the nominal CMY B is the Jacobian matrix given by 2 6 B ¼ 6 @a* 4 @C @b* @C @L* @C @L* @M @a* @M @b* @M @L* @Y (8:17) 3 7 (8:18) @a* 7 @Y 5 @b* @Y The Jacobian matrix B is computed at the nominal CMY using the coarse printer model.19 into Equation 8.20 results in x(k þ 1) ¼ x(k) þ Bu(k) þ Ax(k) þ Bu(k) (8:22) (8:21) where A is a 3 Â 3 identity matrix. Since we are using an integrator V(k) ¼ V(k À 1) þ u(k) substituting Equation 8. The Jacobian matrix can also be measured directly on the printer using techniques described in Ref.448 Control of Color Imaging Systems: Analysis and Design For any nominal CMY input. . the closed-loop state equation becomes x(k þ 1) ¼ (A À BK)x(k) þ BKr (8:23) The gain matrix.

Let the L* of these N gray patches be {di }N . Step 2: Estimate an initial set of N nominal CMY values using the steps described above. the target values lie on the L* axis (a* ¼ b* ¼ 0). We chose the poles to be positive real numbers to avoid oscillations and possible overshoots. extrapolation is used to ﬁnd the TRC values below d1 and above dN. These functions are the three gray-balance TRCs corresponding to cyan. and yellow channels.2].17. If d1 6¼ 0 or dN 6¼ 100.2 0. The pole-placement design is used with three real poles. {Di . respectively. resulting in N control loops running in parallel. Two-Dimensional.6.3 0. P ¼ [0. Ci }N . otherwise process the error to generate V(k) and the new set of N printer input CMY values. For each target.14. and go to Step 5.6. The plot of DE convergence for all the seven control patches as a function of print (iteration) number k is shown in Figure 8.6. The Neugebauer model is used as the surrogate printer and the system is simulated in MATLAB. where i¼1 i¼1 i¼1 Di ¼ 2. Special highlight or shadow corrections described in Section 8. where d1 ¼ 0 and i¼1 dN ¼ 100. and {Di . Mi }N . If the color difference is small. Step 1: Select a set of N uniformly spaced gray-target patches whose gray-balance targets are given. The plot of e(k) for the ﬁrst patch as a function of print number is shown in Figure 8. save the printer input CMYi ¼ [Ci Mi Yi] corresponding to the ith patch. L*a*b*) values. stop the iteration process. Typical values are between 0. we select 7 gray targets with L* of 25–90 in steps of 10 with a* ¼ b* ¼ 0 as gray-balance targets. Step 4: Compute DE color difference between measured L*a*b* with the target L*a*b*. The ﬁnal gray-balanced TRCs for all three separations are shown in Figures 8. We summarize below some key process steps used to generate graybalance TRCs using control approaches for a manual off-line sensor. Step 5: Form the three functions {Di ..55 (100 À di).3 and 0.4 In this example.e.13. Step 3: Print these N patches with CMY values and measure their color (i. . and continue with Step 3.2 may be required prior to generating the ﬁnal gray-balance TRCs. The nominal CMY values for the ith patch is estimated to be CMYnominal ¼ ½255 À 2:55Li * 255 À 2:55Li * 255 À 2:55Li *. A smoothing algorithm can be used to interpolate between points to generate full TRCs for all digital values from 0 to 255. For equivalent neutral gray. The B matrix is estimated at the nominal CMY using numerical differentiation. the closed-loop poles are chosen to be at locations P ¼ [p1 p2 p3].15 through 8.One-Dimensional. magenta. Yi }N . where a pole is chosen for each of the three color patches that is within the unit circle of the complex z-plane. and Spot-Color 449 8. a control loop is needed. Example 8.1 POLE-PLACEMENT DESIGN In pole-placement design.

k.14 * DEab(k) as a function of print number.13 Error e(k) with respect to iteration number. indicating a region where very little toner=ink is deposited on the paper.450 6 4 2 0 e(k) Control of Color Imaging Systems: Analysis and Design L*(k) – Ltarget a*(k) – atarget b*(k) – btarget –2 –4 –6 –8 –10 –12 0 2 4 6 8 Print number k 10 12 FIGURE 8. A shadow is a color or shade with a very high digital value.2 HIGHLIGHT AND SHADOW CORRECTIONS The TRC has highlight and shadow regions. Given a 0–255 digital value range. typically above 230. 8. highlights typically occur between 0 and 20. A highlight is a color or shade with a very low digital value.6. . 12 10 8 ΔE* ab 6 4 2 0 0 2 4 6 k 8 10 12 FIGURE 8.

However.One-Dimensional.18 is an example of a cyan TRC shown with typical highlight and shadow regions. or smoothing as described above works well for most of the range of digital values.15 Cyan gray-balanced TRC for pole-placement design. Determining TRCs using a calibration approach with interpolation. Calibration data for highlights is difﬁcult to produce because the print engine is not capable of reliably depositing a small amount of toner and the sensing of the color in highlight region can be very noisy. particularly those of electrophotographic printers. Figure 8. it does not work well for highlights or shadows in high-end printing systems. Most marking engines can reliably .16 Magenta gray balanced TRC for pole-placement design. 250 200 150 Mout 100 50 0 0 50 100 Min 150 200 250 FIGURE 8. and Spot-Color 250 451 200 150 Cout 100 50 0 0 50 100 Cin 150 200 250 FIGURE 8. Two-Dimensional. curve ﬁtting.

Hence. but not small quantities.18 Cyan gray-balanced TRC with highlight (shaded lower left region) and shadow (shaded upper right region) regions. deposit average quantities of toner. As such. the highlight region of most TRCs has low quality because the calibration data tend to be low quality. introducing a noise factor not as dominant in typical measurements. The contribution of the paper to the sensing measurements is larger in a highlight region.17 Yellow gray balanced TRC for pole-placement design. .452 250 Control of Color Imaging Systems: Analysis and Design 200 150 Yout 100 50 0 0 50 100 Yin 150 200 250 FIGURE 8. FIGURE 8. particularly electrophotographic printers require overcompensation in the highlights due to lack of development.

a value of 1 can be 50 45 40 35 30 Cout 25 20 15 10 5 0 0 5 10 15 Cin 20 25 30 35 Smooth TRC Control patches FIGURE 8. Note that it is possible for a person to specify a color that is more saturated than the physical device can deliver. After gray balance. unwanted contours can occur in dark saturated colors. The active control patch is used during iterations to ﬁnd its Cout value with control-based gray-balance calibration. the lightest printable highlight input values are near Cin ¼ 13. The variation in digital count appears as development on the paper for all active control patches. 8. The passive control patch with the input highlight value of Cin ¼ 1 is a value that is well within the highlight region and is not developable.2. and Spot-Color 453 The shadow regions of most TRCs also have low quality characteristics such as non-smooth development or contours. we show how to produce TRCs that work well for highlights and shadow regions. their corresponding outputs are Cout ¼ 20 and Cout ¼ 46. then.One-Dimensional.19.19 Highlight region of cyan gray-balanced TRC. Here.6. The Cout value for the passive shadow patch is assigned (or user selected). If nothing is done to compensate for the highlights. Two-Dimensional. Nodes are included in Figure 8. full saturation is based on the physical device. In this ﬁgure. In this section. We can have more than one passive control patches. The passive shadow control patch is not directly used during iteration. The passive control patches are at Cin ¼ 0 and Cin ¼ 1 digital counts. which is shown in Figure 8.19 for the four Cin values mentioned above. gray-balance algorithms such as smoothing algorithms. active control patches near the highlight region are at cyan digital input values of Cin ¼ 13 and Cin ¼ 30. Adding more toner does not change the color much at the saturation region because it is nearly saturated. for this example. and the user speciﬁcations. If the TRCs are not smooth and well behaved in the shadow region. In other words. TRCs in the shadow region can be low quality because of the physical device.1 Highlight Corrections We illustrate highlight corrections using the highlight region of the cyan gray-balance TRC. .

the actual value at Cin ¼ 1 was mapped to a lower value. Cin ¼ 0 is another passive control patch as in this case (ideally) no toner is developed.2 Shadow Corrections As discussed above. extrapolation can be used to produce a target Cout highlight value corresponding to the input at Cin ¼ 1. such as linear or polynomial extrapolation. but the printed result is far from certain. If the stability of marking engines is not good for highlights.19. calibration TRCs in the shadow region can also be problematic. Cin ¼ 13 was mapped to 20. of three or more active control patches can also be used to produce adequate results. Control points cannot be joined with linear interpolation in the shadow region without good shadow region smoothing. The maximum desired saturation is 255 because that is the extreme value along the input cyan axis. Using the active control patches at Cin ¼ 13 and Cin ¼ 30. We use another passive control patch at Cin ¼ 255.454 Control of Color Imaging Systems: Analysis and Design present in the image pixel and a print engine can try to print it. After smooth extrapolation was done with these two passive patches (Cin ¼ 0 and Cin ¼ 1). although we chose a Cout ¼ 8 corresponding to Cin ¼ 1.6. then algorithms producing TRCs must extrapolate.20 Cyan gray-balanced TRC with shadow region. Cin ¼ 0 was mapped to 0. 250 200 SmoothTRC Control patches 150 Cout 100 100 150 Cin 200 250 FIGURE 8. then this strategy has to be reconsidered to produce good highlights on a dynamically varying process. In Figure 8. we show the approach to solve shadow region problems for the cyan separation. 8.2. Extrapolation.21. . after extrapolation and smoothing. We also used Cout ¼ 0 corresponding to the Cin ¼ 0 patch. The algorithms produce better results when more passive and active control patches in the neighborhood of the highlight are used to deﬁne the TRCs in the extreme highlight region. we used Cout ¼ 8. In Figure 8. So. When there are no active control patches in that region. and Cin ¼ 30 was mapped to 46. Cin ¼ 1 was mapped to 4.

The ﬁrst step would be to select Ni control grid .One-Dimensional. we show the TRCs with highlight and shadow compensations for magenta.22 through 8. If the ﬁnal TRC were to pass through all of these control patches.3 TWO-DIMENSIONAL PRINTER CALIBRATION STATE-FEEDBACK METHODS WITH The 2-D TRCs for printer calibration can be constructed using a state feedback approach similar to 1-D calibration. The approach described above is used to provide highlight and shadow corrections to all separation TRCs. yellow. we used one passive control patch at Cin ¼ 255 whose output digital value was assigned as Cout ¼ 255 since. Note that the black TRC is produced using channel-wise linearization process. Two-Dimensional.24.21. we show a second active control patch at Cin ¼ 149 (Cout ¼ 230) and the third active control patch at Cin ¼ 132 (Cout ¼ 206). the TRC may deviate from the active control points. In our illustration in Figure 8. The ﬁrst active shadow control patch is at Cin ¼ 166 at which the gray-balance algorithm produces a output digital value. In Figures 8. the maximum desired saturation is 255. and in the course of such a correction. and Spot-Color 250 455 Smooth TRC Control patches 200 150 C out 100 50 0 0 50 100 Cin 150 200 250 FIGURE 8. Cout ¼ 249. In Figure 8.21.6. Cin ¼ 255. and black. which is the maximum amount of toner that the marking engine can deposit. Recall that more than one passive control patch can be selected for shadow corrections. for this example. 8. Hence. the shadow-correction algorithm applied (post-gray-balance control iterations) should create a smooth transition to the ﬁnal passive control point.21 Cyan gray-balanced TRC with highlight and shadow corrections for Cin. and so on. then pixels with Cin greater than or equal to 200 will be saturated to 255. This is not desirable for producing good quality images. Passive control patches may have different output digital values for highlight and shadow regions between separations since each development station may have different constraints with respect to the highlight and shadow developabilities. because it is processed independently during graybalance calibration.

. N1 grid points are placed on the gray axis. The rest of the points in the plane of Figure 8.456 250 Control of Color Imaging Systems: Analysis and Design SmoothTRC Control patches 200 150 Mout 100 50 0 0 50 100 Min 150 200 250 FIGURE 8. N2 points on the cyan primary axis.22 Magenta gray balanced TRC with highlight and shadow corrections for Min. prints along each of the ith axis of Figure 8.9 are computed using a 2-D interpolation=smoothing algorithm.9. N3 along secondary to black. N1 Â N2 Â N3 Â Á Á Á. This same process is repeated for all channels.23 Yellow gray balanced TRC with highlight and shadow corrections for Yin. Once all these grid points are selected. and so on. For example. 250 SmoothTRC Control patches 200 150 Yout 100 50 0 0 50 100 Yin 150 200 250 FIGURE 8. control loops running in parallel are used to control all of them.

By using predictive gray-balance techniques. Typically. It is also desirable to use large-size patches or multiple patches (at different locations on the paper) for the same CMY input to reduce the process and sensor noise by averaging measurements over a large patch or over multiple patches with same CMY input. large N) to better capture the printer nonlinearity.24 Black channel-wise linearized TRC with highlight and shadow corrections for Kin. Hence. Therefore.4 PREDICTIVE GRAY BALANCE In the conventional gray-balance control strategy. For a 20 h long run on a cut sheet digital press. Two-Dimensional. we can reduce the total number of control patches by a factor of two or more that require printing and measurement. it is desirable to use a large number of target patches (i. large N or multiple patches or oversized patches are not allowed.One-Dimensional. Due to cost constraints and competitive pressure. the predictive calibration method becomes especially useful for long print runs (e. 8. In principle. the dimension and the layout of these target N patches are designed to ﬁt a single page to save printing and measurement time as well as material (paper and toner).e. we would end up wasting a total of 80 pages for calibration only. .. Use of lower number of control patches is preferred in practice to save cost. Using the normal control-based technique with state-feedback approach (Section 8.. 20 h) with single=multiple jobs where calibration pages are interspersed within=between print jobs.6. For illustration. N gray patches are used resulting in N control loops running in parallel. we would print 10 patches per each iteration. with two test pages for calibration printed every 30 min. let us assume we decided to use N ¼ 10 control patches for printer gray-balance calibration.6). 10–11 patches with one being pure white to extract media white can be used to construct CMY TRCs and another 10–11 patches for building black K TRC. For a typical printer. any savings in wasted paper is considered useful.g. and Spot-Color 250 457 SmoothTRC Control patches 200 150 Kout 100 50 0 0 50 100 Kin 150 200 250 FIGURE 8.

t2. 52.32. we estimate the new CMY values for the loops where we do not print a patch. . In the ﬁrst iteration. 70. This way. 45. t10 and thus 10 different control loops are running in parallel. (a) Switching Strategy: Assume that there are 10 target patches t1. t10. . . . . then we use a prediction method to estimate their CMY values. From this interpolation. iteration M ¼ set M. 55. . iteration 2 ¼ set 2. we do the reverse that is print and measure for the even set and estimate the CMY values (control actuation) for the odd set. iteration M þ 1 ¼ set 1. . we will print only half the number of patches per each iterations. We divide these 10 loops into two sets: the odd set containing t1. other strategy include ﬁrst printing control patches for all N loops say for two iterations. ). iteration 1 ¼ set 1. for each iterations and for each one of the N loops. We now describe the algorithm for predictive gray balance below. That is. and 15. In the second iteration.8 with cubic interpolation is shown in Figure 8.g.. the predicted (or estimated) values of the cyan components of the ﬁve unmeasured patches corresponding to L* values of 20. From the new values of CMY already computed for other loops using the control algorithm.25.1. we alternate between a set of loops with actual state measurement by printing those control patches and assess which ones do not need printing. and selecting one set at each iteration different from the others in a round-robin fashion (e. we print and measure the patches corresponding to the odd set and use a predictive method to compute the CMY values (control actuation) corresponding to the patches in the even set. In Table 8. . After that print only those that show the largest errors.458 Control of Color Imaging Systems: Analysis and Design Let us assume that there are no replicas required to average out the noise. . and 85. . . . and 85 are estimated respectively as 83. we use the normal iterative approach to update the CMY values as obtained by the control algorithm. t4.8. If we do not print a patch for a control loop. .85. . . (b) Predictive Technique: Ultimately. 37. Using the predictive algorithm. t3. and can work well with predictive algorithm. 25. Other switching strategies include (a) dividing the N loops into M disjoint sets. Let the ﬁve unmeasured patches have the L* target values of 20. let the cyan components of the ﬁve measured and controlled patches plus the two corner points be.9. we need to compute the CMY values.28. in one of the iteration cycles. Linear and nonlinear interpolation=extrapolation or polynomial least-square curve ﬁtting algorithms can be used for estimating the required CMY values. Yet. t9 and the even set containing t2.5. as shown in Table 8. * * we show an example of the DEab convergence response (DEab vs. Since our gray-balance aim is the pure neutral axis. . . . 55. . As an example. iterations) for the case when we print all patches and for the case when we print half of them. 70. If for a given loop. 45. A plot of the data in Table 8. This kind of switching process is repeated till we converge to the true CMY value. . we can save printing of redundant patches. The case described earlier is for M ¼ 2 and each N set containing M loops. we assume a* ¼ b* ¼ 0. we print and measure a patch.

which may be the most color-critical portion of a particular page.One-Dimensional. They can be deﬁned as a ﬁxed set of colors.25 Cyan TRC shown with control and corner points (Note: to convert TRC to 0 to 255 scale use x ¼ (100 À L* )2. ref 8. Two-Dimensional.8 Control Patches Used for Predictive Gray-Balance during One of the Iteration Cycles L* 0 40 50 60 80 90 100 100 C=255 100 60 44 32 20 10 0 100 90 80 70 60 Cout 50 40 30 20 10 0 0 10 20 30 40 50 L* ref 60 70 80 90 100 FIGURE 8.55 and y ¼ Cout Â 2. colors in a customer’s proprietary marked patterns. customer logo colors. or customer-deﬁned colors in the form of an index color table [8].26). Consistent color in these areas may make the difference between success and failure in meeting the colorconsistency requirements. and Spot-Color 459 TABLE 8. which may be Pantone colors (Figure 8. for large background areas. Spot colors are often used.7 SPOT-COLOR CONTROL There are many commercially available packages today that deﬁne speciﬁc colors called spot colors. or can be used.55). .

4870 0.4731 0.5676 2.2986 0.2237 0.0602 0.3009 1.4731 0.2422 0.0386 Loop 1 Loop 2 Loop 3 Control of Color Imaging Systems: Analysis and Design Initial Iteration 1 Iteration 2 Iteration 3 Iteration 4 Iteration 5 9.3067 0.4731 0.2422 0.0041 7.3374 0.8200 1.4351 1.0498 0.5172 2.1106 0.5676 2.7312 0.6253 1.1293 0.0028 0.1235 Loop 1 Loop 2 Loop 3 Initial Iteration 1 Iteration 2 Iteration 3 Iteration 4 Iteration 5 9.5168 0.6345 0.8001 0.9895 3.4473 0.0333 0.3084 1. Iterations for the Case When We Print All Patches and for the Case When We Print Half of Them * DEab (Printing All Patches) Loop 4 8.4473 Loop 9 4.7570 0.0427 0.3054 0.0374 Loop 10 5.0197 0.0970 0.3374 0.3053 0.0063 Loop 5 Loop 6 Loop 7 Loop 8 Loop 9 4.1605 0.9895 3.4299 Loop 5 Loop 6 Loop 7 5.7570 0.2205 0.6187 0.8726 0.7680 0.0957 11.0957 0.4674 0.2398 0.0247 0.0054 7.0957 0.4369 0.4731 7.1856 1.0189 0.4351 0.2169 0.0814 0.5168 8.1856 2.4411 0.3084 1.8200 1.0010 11.0214 5.4473 0.5168 0.460 TABLE 8.8325 0.7570 0.1893 0.7850 8.2099 2.7890 0.4299 0.2522 1.5168 0.0235 0.3067 1.0067 3.4411 0.7131 0.4097 0.0427 0.8275 1.2099 0.2287 0.1252 0.7850 0.9 * DEab vs.0059 0.2422 Loop 8 3.0284 0.4299 0.4473 0.3374 7.4299 0.3374 0.0789 0.4248 1.4473 0.7570 Loop 10 5.4248 1.4870 0.6253 2.7603 0.1352 0.3009 1.5172 4.0193 0.4459 0.4870 .2287 0.0015 * DEab (Printing Half of the Patches) Loop 4 8.2605 0.

26 (See color insert following page 428. process drift.One-Dimensional. yellow. and black (CMYK) recipe using multidimensional proﬁles and store them for future use or manually edit the recipe. and (c) inaccurate recipes (to begin with). and algorithm inaccuracies can lead to error in the recipe with respect to reproducing spot color. Many software packages have the ability to enter or create a cyan. then sensor errors. and (b) top left: 7530. is one of the major . and bottom right: 135. bottom left: cool gray 2. magenta. Two-Dimensional. If the recipes are created on a different print engine. bottom left: 1255. which generally occurs. top right: 7546. top right: 1215. The main problem with this kind of approach is that the CMYK formula that was originally entered or obtained through manual entry may not be very accurate at the time of use due to (a) variations in the print engine state. and bottom right: warm gray 2. (b) variability due to operator error associated with manual adjustments=modiﬁcations to CMYK values.) Examples of Pantone colors reproducible by a digital production printer on a coated stock: (a) top left: 108. and Spot-Color 461 (a) (b) FIGURE 8. Drift in the print-engine state.

the comment %%DocumentCustomColors indicates the use of custom (spot) colors. In a fully automated process that does not involve human intervention. Consequently.462 Control of Color Imaging Systems: Analysis and Design difﬁculties faced by print shop owners regarding spot color reproduction. in the above workﬂow. A typical workﬂow is comprised of the following key steps: (a) Selection and creation of appropriate target values for spot colors can be obtained from a Pantone color table or obtained from a dictionary or created using in-line color sensors. This step can be done off-line. names standardized by Pantone Inc. is often used. Once the user accepts the CMYK recipe. Because imaging can occur over a variety of different printing systems and practiced by a variety of different clients and customers. This kind of automated functionality is called automated spot-color editing [9]. the user still has to select the spot colors manually to create the recipe. The preﬂight step can detect the presence of spot colors in a submitted print job and activate the automated closed-loop control to correct for all spot colors found in the search. Manual adjustment of the CMYK values via trial-and-error process. (c) Acceptance=rejection of the resulting CMYK recipes. can make it very difﬁcult to ﬁnd the best matching recipe. (b) Activation of the automated spot-color editing function at the user interface that determines the CMYK recipes for the target spot colors by executing spot-color closed-loop control algorithms based on sensor measurements. The user then RIPs the image containing the new CMYK values for printing. the images may be checked for spot colors by monitoring the tags automatically in a preﬂight step after they are submitted to the printer but prior to being printed. and then insert the new recipe into a print job prior to printing. initiate running of the closed-loop control algorithm automatically. We show the use of an automated closed-loop control method based on tracking to Pantone-deﬁned or customer-deﬁned target values that are meant to automate the determination of the appropriate CMYK recipe for a particular spot-color target at any point in time using sensed values of the spot color at that time. each interacting with the output color. Adjusting four variables (CMYK) manually.) in the document. the user has no assurance that print shop job can reliably produce colors unless every print job is reviewed by a well-trained expert who is very familiar with all the capabilities of the shop. the creation of the recipe is automatically generated. The spot-color detection routine looks for any standard document convention describing the use of spot colors and their names (as one example. For example. spot-color selection=deﬁnition and insertion of the recipe is still done manually. the spot-color tag=name links the recipe to a particular spot color in the document. the colors may not always be consistent or accurate. which ends up consuming more time and can still lead to unacceptable results. In this workﬂow. if the user’s image is in PostScript1 format. The acceptability criteria may be based on visual inspection of a proof copy with a printed color or simply the mean DE values between the target spot color and the measured spot color. Although. An image processing application names these colors and their CMYK or RGB approximations through the %%CMYKCustomColor .

02 0.’’ Alternatively. So.5 (Color name) %%þ 0 0.0 %%AI8_CreatorVersion: 8 %%For: (John Stanzione) (Spot Color Source) %%Title: (solid to process.0 %%Creator: Adobe Illustrator(R) 8.95 0 (Color name) or %%RGBCustomColor comments in the body of the document. to ﬁnd additional colors.38 (Color name) %%þ 0 0. the image-processing computer RIPs the ﬁle and sends the image to production.62 0 (Color name) %%þ 0 0. As shown in the table. ‘‘%%þ’’ identiﬁes a continuation line. If spot colors are present.0 EPSF-3.0 %AI3_ColorUsage: Black&White %AI3_IncludePlacedImages %AI7_ImageSettings: 1 %%CMYKCustomColor: 0 0 0. The four components (cyan.95 0 (Color name) %%þ 0 0. the spot-color control algorithm is activated .04 0.51 0 (Color name) %%þ 0 0 0.0 8 %%þ procset Adobe_shading_AI8 1.3 0 %%þ procset Adobe_cshow 2.03 1 0. which describes some methods of spot-color snifﬁng from an encapsulated PostScript ﬁle.0 0 %AI5_FileFormat 4.One-Dimensional. is also applicable.02 0. the detection routine examines lines starting with ‘‘%%CMYKCustomColor’’ and lines immediately following that begin with ‘‘%%þ.07 1 0. %%CMYKCustomColor provides an approximation of the custom color (spot color) CMYK values speciﬁed by the color name in parentheses. and Spot-Color 463 TABLE 8.eps) %%CreationDate: (5=24=01) (1:02 PM) %%BoundingBox: 0 0 0 0 %%HiResBoundingBox: 0 0 0 0 %%DocumentProcessColors: %%DocumentSuppliedResources: procset Adobe_level2_AI5 1. If no spot colors are identiﬁed in the image ﬁle. an approach described in Ref.03 1 0.2 0 %%þ procset Adobe_ColorImage_AI6 1. Two-Dimensional. magenta.10 Example of a PostScript Document with Pantone Spot Colors %!PS-Adobe-3.3 0 %%þ procset Adobe_Illustrator_AI5 1. yellow.79 0 (Color name) %%þ 0 0 0. representing the percentage of that process color. [10].81 0 (Color name) %%þ 0 0. Table 8.10 provides an example of a PostScript document with Pantone spot colors and comments describing the Pantone colors.6 (Color name) %%þ 0 0. and black) are speciﬁed as numbers from 0 to 1.

or the print job may be held until the current print job completes. obviously. In essence.6.5. time savings.7. DE2000 is one of the preferred [11] methods since it can map the out-of-gamut colors perceptually close to the gamut boundary. 8. The advantages of running spot-color control automatically include reduction of long-term engine drift effects. An example of a tricolor gamut class is comprised of four gamut classes . map the colors in reference to a point in the surface of the printer’s gamut.4. 8. and reduced operator intervention. The overall color printer gamut is represented as a composite of the gamut subclasses wherein each gamut subclass is comprised of a subset of CMYK. The spot-color control may run while another job is printing. A color that is wrongly classiﬁed as outside the gamut will be handled by the gamut mapping algorithm. A spot-color control algorithm is comprised of the of following key steps: (a) Determining whether or not the spot-color targets are inside the printer gamut (b) Mapping out-of-gamut colors to printable colors using appropriate gamut mapping algorithms and determining the new target L*a*b* values (c) Selecting appropriate GCR (or gamut classes) for the target L*a*b* values (d) Applying closed-loop control algorithm for the GCR constraints set in Step c (e) Selecting best CMYK recipe out of multiple iteration steps These steps are described in detail next. reduced operator error. this could have been avoided since there was no need to map the original color. The ray-based control model described in Section 7. spot-color stability. Selected spot-color targets are assigned to one of the gamut subclasses for efﬁciently calculating the CMYK recipe for a given spotcolor target. A tricolor GCR is preferred for spot colors due to the advantages described in Section 7. but is conceivable for digital printing systems of the future. Any of the gamut clipping or gamut compression methods can be used to map out-of-gamut colors.3. Thus. this kind of automation would decrease the time needed to improve the accuracy and repeatability of spot colors in print jobs. this mapping algorithm will.7.1 can be used to determine whether the spot-color targets are inside or outside the gamut surface.464 Control of Color Imaging Systems: Analysis and Design automatically. such a system is still not commercially available.1 GAMUT MAPPING FOR SPOT-COLOR CONTROL One of the key components in ﬁnding a CMYK recipe for a given spot color is determining whether the target L*a*b* values are either inside=on-boundary or outside of a printer’s gamut.2 GAMUT CLASSES The basic algorithm requires the use of GCR-constrained inversion for the target L*a*b* values of the given spot color. The consequence of this action may lead to the reproduction of a color that is slightly different from the original one. in turn. Colors located very near or on the gamut’s boundary could be mistakenly reported as outside the gamut by algorithms that are not sufﬁciently accurate. Currently.

(CMY ! L*a*b*. This can be a forward printer model obtained via experimentation on a color printer or a mathematical .27 shows an additional region classiﬁed as the neutral zone. Figure 8. and CMK ! L*a*b*) and a neutral zone. MYK ! L*a*b*. and (d) CMK ! L*a*b* and CMY ! L*a*b* gamuts (alternate view shown with neutral zone).27 (a) CYK ! L*a*b* and MYK ! L*a*b* gamuts.One-Dimensional. and CMK ! L*a*b*. Figure 8. a three-input three-output MIMO control algorithm can be implemented to achieve improved spot-color accuracy. The overall gamut is the union of all of them. CMY ! L*a*b*. (b) CMK ! L*a*b* and CMY ! L*a*b* gamuts. and Spot-Color 465 FIGURE 8.27a through d shows in L*a*b* space for the four gamut classes. These gamut classes represent the color reproduction capability of a printer obtained by considering three separations at a time. This approach reduces the dimensionality of the four color process to three color groups. (c) CYK ! L*a*b* and MYK ! L*a*b* gamuts (alternate view shown with neutral zone). respectively. Two-Dimensional. CYK ! L*a*b*. As a result of this reduction. MYK ! L*a*b*. CYK ! L*a*b*. In plots c and d.

and MYK gamuts is low. this strategy has two components: (1) minimum black strategy and (2) maximum black strategy. geometrically looks like a cylinder of circular cross section in 3-D space when the cylinder is * speciﬁed with DEab as a radius from the neutral axis. In summary. (c) more room for the . at a high level. it reduces the dimensionality of the four-color process to three-color groups.. we use black everywhere (i.4. [12].) Top view of CMK. Whenever there is overlap with the CMY gamut class. In speciﬁc terms. Whereas the overlap region of the CMK. we use the CMY gamut class wherever possible. this approach provides (a) improved spot-color accuracy with one of the separations always held to zero. CYK.28 (See color insert following page 428. For minimum black strategy. as shown in Figure 8. and MYK gamuts can cause problems while assigning spot colors to appropriate gamut classes.28 shows the CMK. MYK. Figure 8. This zone. In some cases.e. and CYK gamuts along the chromatic axes when viewed from the top of the gamut. CYK. The use of tricolor gamut classiﬁcation algorithm removes the degeneracy while ﬁnding the CMYK recipes. MYK. To reduce the overlap with the CMY gamut class and to preserve the appearance of images. A classiﬁcation algorithm is shown in Ref. this neutral zone may not be needed. no CMY gamut is used).466 Control of Color Imaging Systems: Analysis and Design FIGURE 8. the neutral zone (region) is deﬁned using the DE metrics. For maximum black strategy. model of the type described in Section 7. The overlap of the CMY gamut with the CMK. (b) improved toner usage for high area-coverage spot-color printing by identifying the most toner efﬁcient CMYK values for each spot color. It is much less cylindrical when the radius is speciﬁed in DE2000 or perceptual space. CYK gamuts and neutral zone (CMY gamut removed from the ﬁgure).27c and d. There is a high degree of overlap between the CMY gamut with other classes that include the K separation.

7. 8.2) algorithms can be used to ﬁnd the estimated tricolor recipe associated with each mapped or in-gamut spot-color target using the printer model (analytical or empirical. the CMY values may go to outside the range 0 or 255). u(k) ¼ þKe(k). The approach is shown in Figure 8. .29 is stable. is derived based on the pole values speciﬁed such that the closed loop shown in Figure 8. The recipe obtained from the ICI or moving-matrix algorithms are used as nominal CMYK values while working with the iterative feedback loop corresponding to the gamut class under use. as described in Figure 7. which will place the eigenvalues of the closed-loop system inside the unit circle. To further reﬁne the accuracy. For out-of-gamut colors.4 CONTROL ALGORITHM WITH INK LIMITS Often the tricolor gamut class does not give full gamut coverage in the dark part of the gamut since those colors cannot be produced with only three separations. for the spot color at or near the gamut boundary is higher than the previous one. During iterations.41. but the same method applies to the rest of the gamut classes.4. choice of gamut class not only depends on the color. a three-input three-output MIMO state-feedback controller described in Section 7.5. we know which of the three color separations to use to ﬁnd the recipe for a given spot color. then a best CMY selection algorithm (see Figure 7. often a four-color gamut class is introduced. This would require more complex control algorithms. if the error. K. Iterations can also be carried out directly on the printer with an in-line color sensor. Constraints for ink limits can be applied during iteration while ﬁnding the recipe automatically.32) can be used because the CMY values near the boundary may reach saturation (i. To reproduce spot colors accurately with full gamut coverage. Accuracy of this recipe depends on (a) accuracy of the printer model and (b) accuracy of the inversion algorithm. A GCR-constrained 4-to-3 control-based inversion. Ink limiting constraints are often required in a four-color system due to limitation on the fuser.7.2 is used. can be used.3. and Spot-Color 467 controller regarding the actuators when the spot colors under consideration are common to two or more gamut classes. The gain matrix. Thus. 8.One-Dimensional. The ICI (Section 6. This is achieved by assigning pole values in the range [0. obtained using experimental data) associated with each gamut class. The control law can be designed using MIMO state-feedback methods or LQR or model predictive control [13] approaches.29 for CMY gamut class). but also depends on the gamut mapping strategy used to map the colors to the printer’s gamut.3 CONTROL ALGORITHM Once a gamut class is selected. where e(k) is the error between the target and the measured L*a*b* at iteration k. and (d) a unique solution for each spot color once a gamut class has been chosen.2.29 for CMY gamut class.e. 1). e (shown in Figure 8.3) or moving-matrix (Section 6.. Two-Dimensional.

Closed-loop control algorithm with a gain matrix and the integrator as controller shown for spot-color control algorithm using CMY .468 Feedback controller x(k) Initial/nominal CMY values obtained using ICI algorithm Reference r K e(k) u(k) Integrator + + Spot color L*a*b* Printer model for CMY gamut class (or input–output data) Measured L*a*b* x(k + 1) = Ax(k) + Bu(k) Mapped L*a*b* for out-of-gamut colors Control of Color Imaging Systems: Analysis and Design FIGURE 8.29 gamut class.

89 27. i ¼ 0.5 7.5 0. 8.11 * DEab Difference between the Output Patches and Paper Y * DEab 0 0 25.13 0.08 .19 0.72 7.08 0.08 0. . (a) Find the cyan.One-Dimensional. 8.12.11 0.98 3.32 0.45 a* 0. and Spot-Color 469 PROBLEMS 8.3 102 12.5 20 153 29 178.05 0.09 b* 0.5 229.5 Use Neugebauer model.33 36.5i.87 0.3 Ten gray patches are printed on a digital printer and their L*a*b* values are measured. 2.8 51 3.12 0.12 Gray Patches: CMY!L*a*b* C 0 25 51 76 102 127 153 178 204 229 255 M 0 25 51 76 102 127 153 178 204 229 255 Y 0 25 51 76 102 127 153 178 204 229 255 L* 92. (b) What would the resulting measured L*a*b* values be if the patches are printed before and after calibration? * 8.22 55.61 47. TABLE 8. drive a closed-form expression for the sensitivity matrix B at a nominal CMY value.07 0.7 76.09 0.1 Ten yellow patches with digital input of Y ¼ 25.08 0.78 18.19 0.13 0.09 0. 1. Find the yellow TRC for channel-independent linearization. Linearize the overall system and check the stability of the system. 10 and * C ¼ M ¼ 0 are printed on a digital printer. Two-Dimensional.3 76.11. . The DEab difference between the output patches and paper are tabulated in Table 8.08 0.9 67.4 Design a control loop for 1-D DEab from paper linearization using state feedback * where the plant is modeled as combination of printer and DEab sensor.5 39.06 0. and yellow gray-balanced TRCs for this printer.5 204 51. . . magenta.5 65 255 80 TABLE 8.2 (a) Write a general MATLAB code to ﬁnd the inverse of function f(x) graphically from data given by a LUT.8 127.12 0 0.09 0. 8.5 88. The resulting data is shown in Table 8. (b) Write a general MATLAB code to ﬁnd the inverse of function f(x) numerically from data given by a LUT.

L.K.355. Mestha.7 Perturb the sensitivity matrix with reasonable DB and show the robustness of spot-color recipe obtained in Problem 8. 2006. 2008.K. J. Calibrating.P. L. Li.143.456. Hoffmann. Balasubramanian. Mestha.6 Use the sensitivity matrix B obtained in Problem 8. 11–13.K. Mestha. NIP21: Proceedings of IS&T’s International Congress on Digital Printing Technologies. T. P. Crean. and T. G.D. On-line calibration system for a dynamically varying color marking device. A method for classifying a printer gamut into subgamuts for improved spot color accuracy. Fisher. Two-dimensional calibration architectures for color devices. Y. Sharma. Sensitivity matrix determination for adaptive color control. Dianat. L. MYK.K. Sep. 2005. and E. 4. Dec.K.470 Control of Color Imaging Systems: Analysis and Design 8. L. US Patent 6. Use the pole-placement algorithm and three different combinations of poles. 499–504. Bala. Spot color control system and method. L. 9. 2002.J. 21. J.S. P. Mestha and S. S. 2008.N. Feb. M. 18–23. and L. Gray balance control loop for digital color printing systems. 8.752. 2008. Printing and Prooﬁng to the G7 Method. E. Sept. Ringness. MD. 13. Furst. Baltimore.K. Maltz. US Patent 7. T. and S. 2007. System and method for automated spot color editor. 10. Wang. Love. 2006. K. Sep. . Mestha. Spot color controls and method. 505–510. Accurate printing of proprietary mark patterns and colors. 8.E. Mestha. M. US Patent 7. 11. 21. 19. 2008. US Patent Application. Bala. V4. and M. 8. Dianat. Gil and L.R.R. Hoffmann. Feb. 2005.F.K. Mestha. 8. 2-D and 3-D printer calibration algorithms with printer drift.6 for CMY gamut class.F.F. Gil. Crean. 7. Harrington. 12. US Patent Application. P.752. Sep. Dec. Y.H. 18–23. R. Gil. Malz.6 and 8. 20080043264. Proceedings of 2008 American Control Conference. A.A.A. Viturro. Feb. L. Feb. and M.7 for CYK.R. US Patent Application. 3.K. Baltimore. US Patent Application. Bares. 2.5 to simulate the in-gamut spot-color convergence when the spot color is inside the CMY gamut class. L. Mestha. Rolleston. US Patent 7. A. 20080043271. Attorney Docket 20071048. Hancock. R.L. pp. Hoffmann. and CMK gamut classes by choosing in-gamut colors in the respective gamut classes. Mihalyov. S.J. E.307. Sim and P.A.J. and M. 5. R.E. L. 11.Y.8 Repeat Problems 8. Apr. Mar.E. M. A. Attorney Docket 20061087. Jun.S. Seattle. MD.395. pp.K.K. Comparison of 1-D.A. NIP21: Proceedings of IS&T’s International Congress on Digital Printing Technologies. Dianat.R. A full GRACoL technical speciﬁcation document. Mestha. Jackson. Method for separating colors of encapsulated postscript images. 2005. On coordination and stabilization of two xerographic printers. J. REFERENCES 1. US Patent Application 20080043264. Jackson.K. P. 2008.E. Mestha. 6. Van de Capelle. Gurram. 2007. and R. WA.110. Wang.

general wear. These models are used to design a time hierarchical control system for controlling all the processes that affect the images printed on paper. 471 . Unique strategies and methods to overcome the adverse effects of loop interactions and avoid undesirable glitches in the controller performance are also presented. a time hierarchy is preferred for simplifying the implementation complexity of such a system. These factors include environmental conditions (temperature. Control loops were single-input single-output (SISO) although they were implemented on a coupled multiple-input multiple-output (MIMO) system. we describe how the system is represented in a system theoretic form. use patterns. the type of media.and highdensity regions of the tone curve (e. aging of the components. and tone reproduction curves (TRCs) of the individual primaries by adjusting various internal process and image actuators that operate at varying frequency while making prints. for example. Xerox product. First. The control algorithms for set-point tracking were done through ‘‘on–off’’ rules. etc. and stability of the internal states of the system.. variations from original models used in initialization.1 INTRODUCTION The output of digital printers drifts over time or deviates from the predetermined optimum standards due to a variety of factors.). Many earlier color products only used solid area control for low. important internal parameters (states of the machine) are controlled by applying feedback [1]. variations in media.9 Internal Process Controls 9. proportional. 5775). Today’s digital production or entry production systems have become more complex and require a multivariable. often called a state-space model. integral control was used in the SISO control loop. when certain conditions are met. design of the key process control loops and their effects on the ﬁnal print quality are described in detail. otherwise not. After that. This corresponds to one-point TRC control without any hierarchy.g. dispense the toner. To achieve predictable print quality time after time. On some occasions. In this chapter. solid area development. etc. modular design approach to implement stable feedback systems that can deal with a wide range of process parameters and ensure the accuracy. theoretical methods are developed for analyzing the control performance of processes within the xerographic printing system. relative humidity [RH]. consistency. These loops maintain background.

there are well established results available in the literature for the control of such systems using state or output feedback control. For typical control applications.e.. and D are constant). we are especially interested in the linear version of the system given by _ x ¼ A(t)x þ B(t)u y ¼ C(t)x þ D(t)u or x(k þ 1) ¼ A(k)x(k) þ B(k)u(k) y(k) ¼ C(k)x(k) þ D(k)u(k) (9:2b) (9:2a) (9:1b) (9:1a) If we can express the system in this form. If the linear system is also time invariant (i. This version can be written in continuous time as _ x ¼ Ax þ Bu y ¼ Cx þ Du (9:3a) . k) where x is the system state u is the input actuator y is the output quantity t and k represent continuous and discrete-time instants. respectively In fact. B. u(t). C. this is not enough for control. and is something quite innovative at the time this book was written. u(k). we need a representation of the system in the form of a vector differential equation in continuous time _ x ¼ g(x(t). as an analytical MIMO system is very important for applications. u(k). when the matrices A. k) y(k) ¼ h(x(k).2 PROCESS CONTROL MODELS—A GENERAL CONTROL VIEW The compact characterization of a printer. then simple design techniques are readily available for powerful controllers and there are simple criteria for controllability and observability. described in Chapter 10. t) or the discrete time version x(k þ 1) ¼ g(x(k). However. t) y ¼ h(x(t).472 Control of Color Imaging Systems: Analysis and Design 9. u(t).

C @ . Let x be the state vector which is also the output. C . . A . . @ . through the Jacobian matrix of the system which is by deﬁnition the derivative of the outputs with respect to the inputs. The answer is simple. C .Internal Process Controls 473 or in discrete domain as x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) þ Du(k) (9:3b) The question is how to convert the nonlinear static MIMO representation modeled in Chapter 10 to a suitable linear time-invariant (LTI) form discussed above. fm (u) 0 1 u1 B u2 C B C u(t) ¼ B . so y ¼ x. un The Jacobian matrix can then be expressed as the following m Â n matrix: 0 @f (u) @f (u) 1 1 1 Á Á Á @f1 (u) @u1 @u2 @un B C B @f2 (u) @f2 (u) Á Á Á @f2 (u) C B @u1 @u2 @un C C J(u): ¼ B B . @fm (u) @u1 @fm (u) @u2 df (u(t)) du (9:4) 0 (9:5) (9:6) (9:7) ÁÁÁ @fm (u) @un Applying the chain rule of calculus we obtain _ x¼ @f (u(t)) du _ ¼ J(u(t))u @u dt (9:8a) The discrete-time case is similar with t replaced by k x(k þ 1) ¼ x(k) þ J(u(k))Du(k) (9:8b) . . . C @ . and let u the input vector to the nonlinear MIMO system. For the case x ¼ f(u). the Jacobian is given by J(t) ¼ where f(u) is an m Â 1 vector u(t) is an n Â 1 vector that is 1 f1 (u) B f2 (u) C B C f (u) ¼ B . A . A . .. B .

8b represents a typical linearization scheme for a nonlinear discrete-time dynamic system around a time-varying point. B(k) ¼ J(u(k)). It is vital to recognize that for this approximation to hold true.12 is a generalized vector linear time-varying representation of a printing system with actuators (inputs) and outputs.474 Control of Color Imaging Systems: Analysis and Design where Du(k) ¼ u(k þ 1) À u(k) (9:8c) In the rest of this chapter. we get x(k þ 1) ¼ x(k) þ J(u(k))v(k) which is a linear time-varying state-space model of the form x(k þ 1) ¼ A(k)x(k) þ B(k)v(k) y(k) ¼ C(k)x(k) þ D(k)v(k) A(k) ¼ I. First deﬁne v(k): ¼ Du(k) ¼ u(k þ 1) À u(k) (9:9) which is mathematically equivalent to the use of a discrete-time integrator for the calculation of u(k) from v(k) (since we have access to the input u(k) for control). The Equation 9. The size of the deviation depends on the local nonlinearity of the model. For the special case where the system is linearized about a ﬁxed. namely u0 (which gives in turn a nominal value of the output. with respect to the reference point of linearization.8b becomes x(k þ 1) ¼ x0 þ JDu(k) (9:13) . which has to be reevaluated at each step. Thus. if the entire sensing-processing-actuation cycle is implemented within the printed page.3b. that is u(k þ 1) ¼ u(k) þ v(k) Substituting into Equation 9. The index k can also refer to the sample measurements within a single page. where k refers to index of the print. Equation 9. This model is trivial to put into the state-space form of Equation 9. we will focus on the discrete-time domain. the only time-varying component is the system’s Jacobian matrix. C(k) ¼ I. namely x0) Equation 9.8b. D(k) ¼ 0 (9:11) (9:10) (9:12) where. x(k) and v(k) must have relatively small deviations between consecutive time instants. time-invariant nominal value of the actuators.

we obtain x(k) ¼ x0 þ JDu(k À 1) (9:19) (9:18) (9:17) (9:16) Solving Equation 9. this equation is still not in the typical state-space form of Equation 9.16 into Equation 9.1 below for an example system from Chapter 8. From the analysis above. we have x(k þ 1) ¼ Ax(k) þ Bv(k) y(k) ¼ Cx(k) þ Dv(k) (9:21) (9:20) where A ¼ I.19 for x0 and then substituting it into Equation 9. These are the same reasons why it would . Most of the difﬁculty stems from the fact that the process models for each separation are not independent and additionally because the color model employed presents a very complex dependency on the toner masses. and D ¼ 0.18 we get x(k þ 1) ¼ x(k) þ Jv(k) which is of the desired form. B ¼ J.Internal Process Controls 475 where matrix J is the Jacobian computed at u ¼ u0 J ¼ J ð u0 Þ and Du(k) ¼ u(k þ 1) À u0 (9:15) (9:14) However. we observe that the use of a discrete integrator results in a very neat and elegant linear state-space representation.21 is fairly simple from a theoretical perspective.13. Putting everything together. C ¼ I. Although Equation 9. To get an equation of the form of Equation 9. it is difﬁcult to apply in practice.15) u(k þ 1) ¼ u(k) þ v(k) Substituting Equation 9. which is the most desired form. we get x(k þ 1) ¼ x0 þ J(Du(k À 1) þ v(k)) Replacing k þ 1 with k in Equation 9. This is so because the models describing the printing system are very complicated and have some very challenging nonlinearities. Matrix I is a 3 Â 3 identity matrix. This is schematically presented in Figure 9. we use a discrete integrator for Du(k) and deﬁne Du(k) ¼ Du(k À 1) þ v(k) or equivalently (by the deﬁnition of Du(k) in Equation 9. Note that as shown v(k) is deﬁned as v(k) ¼ ke(k).3b.3b.13.

The magnitude of each of the entries in the Jacobian is an index of robustness of the system. . for example. This process is not trivial. a lookup table (LUT).476 Control of Color Imaging Systems: Analysis and Design Nominal CMY values Aims L*a*b* – Printer and sensor + e K v Integrator u + x x(k + 1) = Ax(k) + Bv(k) Measured L*a*b* FIGURE 9. 2. is easily avoided by making the assumption that F ¼ 0 or F ¼ 1 with the proper control of Va in the transfer model (Table 10. it is possible to ﬁnd the Jacobian using a computer application like MATLAB symbolic toolbox. What are the limits of uncertainty over its operational boundary? Can we design the feedback system robust enough by considering the bounds on the Jacobian so that the actuator latitudes are preserved during a variety of operational scenarios with different media and images acting as disturbances to the control system? Section 9.3 includes a discussion of the Jacobian for different process control hierarchies. however. However. Robustness: Robustness with respect to measurement noise is a very important property since it provides insight regarding the behavior of the system in the presence of noisy measurements. which in turn means that the system is no longer full rank. This is quite typical in practice and is equivalent to ignoring the transfer model from the system and substituting it by. which uses an algorithm to minimize the minimum absolute singular value or. all designed for increased performance. The sensitivity analysis is very important. A heuristic approach has been followed.8). to see for what values this quantity approaches 0. the Jacobian provides invaluable information about two important properties of the system as explained below: 1. it is shown in the following section that the LTI system approximation is controllable if and only if the Jacobian matrix is full rank.1 MIMO state-space model with the integrator used in spot color control model for CMY separations. This difﬁculty. The Jacobian also helps answer questions like the two that follow. be very difﬁcult to derive the Jacobian theoretically. equivalently. Controllability: Because of the special form of the state-space model. because it opens the way toward the design of sophisticated real-time state-based feedback controllers and the use of model predictive control (MPC). but an algorithm can be implemented that derives the Jacobian pretty fast ($1 s). Apart from that.

An abstraction of the architecture for levels 1. are several layers higher than the subsystem level controls. In a control hierarchy. Control functions that require ‘‘human=operator-in-the loop. not only due to the use of multiple primaries. and algorithm properties. Time hierarchy comes from the ‘‘reduction of complexity’’ rule used to design complex control systems. 3.g.. levels 1. controlling a group of subsystem variables at a higher rate. Each layer in the architecture is characterized by the nature of its sensor input.’’ etc. The algorithms of these lower level controls must be both noise immune and able to respond rapidly to changes in set points since their set points will be altered by the higher level algorithms. 2. level 3 the image control for each separation tone adjustments (e.3] that is conceptually very similar to other hierarchical control systems [4]. which is a multilayer belt or drum that retains charge in the dark but discharges when .. It operates at the subsystem level to control subsystem parameters directly. 1-D tone reproduction control).2a for an imageon-image (IOI) printing system with a belt photoconductor. 2. Level 1 includes the lower level subsystem controls such as the ‘‘charge control. An important feature of the level 1 controller is that there is a simple and direct relationship between the sensed and controlled parameters. The higher level controls coordinate commands to subsystems at a much lower rate.Internal Process Controls 477 9. and 4 controls are used to describe the time hierarchy of process controls. ‘‘charge and development’’ systems). The lower level controllers deliver simpler view to higher level controls.’’ ‘‘toner concentration control. the actuations and the sensed parameters are coupled by a single process step. Both its sensing and actuations occur locally. which transforms the printing system to many simpler subsystems while preserving the overall performance goals. but also due to high print quality requirements for color due to the eye’s sensitivity to small color variations. At this level. The control functions managing the job scheduling and managing set points based on media attributes (on a sheet-by-sheet basis) are done by constraint-based schedulers that are executed at a different level. It is important to understand the principal subsystems involved in the design of the level 1 controller. as in the drum photoconductor system. the lower level controllers run faster than the higher level loops. redirecting jobs to an available printer when another fails. In Xerox.g.3 TIME HIERARCHICAL PROCESS CONTROL LOOPS The controls required for color printers are signiﬁcantly more complex as compared with those required for black and white printers. level 2 the controls between subsystems (e. 9. Each controller sees the controllers below it as a virtual body from which it gets information and sends commands... a multilevel modular control architecture was proposed [2.4 LEVEL 1 ELECTROSTATIC CONTROL SYSTEM The level 1 controller is closely coupled to its related subsystem. Given the traditional difﬁculties associated with the control of color. 2-D LUT and 3-D proﬁles) to minimize the interactions between colorants that cause color shift in the output. The xerographic process is centered on the photoconductor.’’ for example. actuator output. Similar abstraction is applicable to a typical printing system that is non-IOI. and level 4 the image control between multiple separation tone adjustments (e. and 3 is presented for process control loops schematically in Figure 9.g.

Vg. (b) A typical xerographic print engine with a drum photoconductor showing principal subsystems. X (Table 10. The sensor used to measure the voltage levels on the photosensitive surface is a nonconducting electrostatic voltmeter (ESV) or electrometer. for both unexposed (charged) and exposed (discharged) regions. The ﬁrst subsystem begins with an uncharged photoconductor and ﬁnishes with the exposed customer page (the mirror image. exposed to light. The actuators that are normally used for controlling the photoconductor surface potential in this process are the grid voltage. [8]. maintains the charge on a moving photoconductor surface at desired set points. and measures the voltage level of the photoconductor surface as it traverses under an ESV probe before and after the . the level 1 controller. The electrometer is generally rigidly secured to the printer. and the exposure intensity.3) [5–7]. also called the latent image) expressed as a pattern of discharged dots. sensor examples are available in Ref. in its simple form. adjacent to the moving photoconductor surface. Therefore.478 Control of Color Imaging Systems: Analysis and Design PE image path CMYK TRCs Electrostatics Level 3 Level 2 Level 1 Development system Targets Printer with sensors Voltage sensor TC Loop Belt photoconductor DMA sensor (a) Charge Expose Clean Noncontacting electrostatic voltmeter (ESV) Fuse Develop Media transport (b) Develop Optical sensor FIGURE 9.2 (a) Abstract representation of multilevel control architecture.

The exposed patch area voltage level. Vl. between the latent images as illustrated in Figure 9. The surface voltage is a measure of the density of the charge on the photoconductor. which is related to the quality of the printed output. A DC measurement circuit is combined with the ESV circuit for providing an output that can be read by a conventional test meter or used as input to the feedback control computer. There is typically a routine within the operating system of the printer to periodically create test areas (called patches) [9. h Control patches Photoconductor Latent image area Latent image area Latent image area FIGURE 9. A typical ESV is controlled by a switching arrangement that provides the measuring condition in which charge is induced on a probe electrode corresponding to the sensed voltage level on the photoconductor. The desired goals for the feedback system is to maintain the charge (and hence voltages Vh and Vl) on a moving photoconductor surface to desired set points (V T and VlT respectively). and an unexposed patch area voltage level. such as the inter document zone (IDZ).3.3 IDZ patches (shows typical patches used for measuring exposed and unexposed areas of the CMYK separations using four ESVs).4 Block diagram representation of the level 1 electrostatic control loop. Ugo – VT + h VT l Eh E1 + – k11 k21 k12 k22 Controller + Integrator u1 + Integrator + ug ul + Ug Electrostatic charging and exposure system Vh Sensor values Vl u2 + + Ul Ulo Target values FIGURE 9.4.10] at predetermined locations on the photoconductor by deliberately causing the exposure system to charge or discharge as necessary for measurement with the ESV. the grid voltage. The induced charge is proportional to the sum of the internal capacitance of the probe and its associated circuitry relative to the probe-to-measured surface capacitance. Ug (also denoted by Vg for later use) . Vh. A block diagram representation of the level 1 control system is shown in Figure 9.Internal Process Controls 479 exposure. are shown in the diagram with the actuators. To avoid negative impact on productivity. color patches are created in small unobtrusive areas. Ul (also denoted by X for later use). and the exposure intensity. .

X ¼ Ul .5a at the point marked ‘‘x’’ at {Vgo.Vlo}. ul}. Points on these ﬁgures marked by ‘‘x’’ indicate a nominal operating point (chosen at random to illustrate the approach).5b is basically a laser power curve. a photo induced discharge curve (PIDC) at the fully charged photoconductor potential. Vho}. Let b11 be the slope of the curve in Figure 9. The expressions for the deviations in photoreceptor voltages.5c that has the coordinates {Vgo.4) be the deviation about Xlo. caused by the electrostatic controller. if the laser power is set equal to Xlo.5a through c with notations V go ¼ U go .5b that has the coordinates {Xlo. as follows: DVh ¼ b11 DUg DVl ¼ b21 DUg þ b22 DUl (9:22) –450 –500 Vh (volts) –550 Vho X –600 –650 Vgo = Ugo –700 –700 –680 –660 –640 –620 –600 –580 –560 –540 –520 –500 (a) Vg = Ug (volts) FIGURE 9. DVh and DVl. Vho is the voltage on the unexposed photoreceptor.4) be the deviation around Vgo which would be generated by the controller when the charging control loop is closed. DUl} {ug. for the grid voltage set to Vgo. Let DUg (used synonymously with ug in Figure 9. Vlo}.’’ When there is no feedback. Xlo ¼ Ulo . Let b22 be the slope of the curve at point ‘‘x’’ in Figure 9. respectively.’’ . Vho and Vlo are voltages at the nominal operating point ‘‘x. Let DUl (used synonymously with ul in Figure 9.5a and c plot photoconductor potentials as a function of the grid voltage for the unexposed and exposed regions. V g ¼ U g .480 Control of Color Imaging Systems: Analysis and Design The model-based curves for a typical electrostatic system are shown at a nominal operating point in Figure 9. With the grid voltage remaining at Vgo.5 (a) Vg(Ug)ÀVh curve gives slope b11 at point ‘‘x. Figure 9. let b21 be the slope of the curve at point ‘‘x’’ in Figure 9.5b). can be written in terms of the small signal deviations {DUg. then the photoconductor will be exposed to Vlo volts (shown in Figure 9. Figure 9. Similarly.

5 (continued) (b) X (Ul)ÀV1 curve for nominal Vg(Ug) gives slope b21 at point ‘‘x.22 for the nominal operating point is shown in block diagram form in Figure 9.’’ Note that we assume small deviations so we can ignore all the second and higher order terms in Equation 9. The actuator signals before and .6.22 so that the system equations are simple and linear [11]. The linear electrostatic system described by Equation 9.’’ (c) Vg (Ug )ÀVl curve for nominal X(Ul) gives slope b22 at point ‘‘x.Internal Process Controls 0 481 –50 Vlo –100 V1 (volts) X –150 –200 –250 Xlo = Ulo –300 2 (b) 4 6 8 10 X = Ul (ergs/cm2) 12 14 16 –50 –60 –70 –80 Vlo V1 (volts) –90 –100 –110 –120 –130 –140 X Vgo = Ugo –150 –700 –680 –660 –640 –620 –600 –580 –560 –540 –520 –500 (c) Vg = Ug (volts) FIGURE 9.

the exposed voltage. B is the 2 Â 2 Jacobian matrix given by B¼ b11 b21 0 b22 ! (9:23) The state vector x(k). As the photoconductor charge is reduced in magnitude or increased from more negative to positive value (see Figure 9. the slope.6 Block diagram representation of linear electrostatic system (Equation 9.1 Test the following electrostatic control system for controllability. is not affected by the laser intensity. is equal to zero.21) by introducing the integrator.7 at nominal X(Ul) ¼ 8 ergs=cm2). it can result in loss of controllability as described above. where A ¼ I is the 2 Â 2 identity matrix. x(k) ¼ Vl ! ug ¼ DUg v(k) ¼ . when b22 ¼ 0 the states x(k) are not fully controllable using only the two electrostatic actuators. and the output vector y(k) are given by ! Vh . respectively. Example 9. after the integrator are {u1. y(k) ¼ x(k) ul ¼ DUl (9:24) It is clear from Figure 9. Xlo ¼ 8 ergs=cm2 }.6 that if the slope. If the slope b22 is too small. b22. decreases rapidly. ul}. The system is at the nominal operating point {Vgo ¼ À600V. Vl. u2} and {ug. x(k þ 1) ¼ 1 0 1 y(k) ¼ 0 ! ! 0 0:9798 0 x(k) þ v(k) 1 0:3315 11:1026 ! 0 x(k) 1 . In other words. b22. The control system is modeled in state-space form (Equation 9.482 Control of Color Imaging Systems: Analysis and Design Vgo= Ugo Ug = ug + b11 b21 b22 + Vl = Vlo + Vl Xlo = Ulo Vh = Vho + Vh Ul = ul + FIGURE 9. control vector v(k).22) shown with nominal actuator values.

À500 V.1 ELECTROSTATIC CONTROLLER DESIGN From the theory of linear systems. this condition is trivial and simply requires that matrix B is full rank.21. since the output is equal to the state vectors (C ¼ I). we need to ﬁnd a condition for controllability. Since A ¼ I. If this is satisﬁed. it is known that two of the most important properties to examine before designing a controller are controllability and observability (see Chapters 4 and 5). Since A is identity matrix.7 X(Ul)–Vl curves for Vg(Ug) ¼ [À300 V (upper curve). then state-feedback control (equivalent to output-feedback control in this case) is feasible and arbitrary pole placement of the closed-loop system can be carried out via many techniques through the appropriate selection of the feedback gain matrix K (see Figure 9. by deﬁnition.1). This is an important condition for control to be successful and is usually satisﬁed everywhere except at the boundary limits of the actuators. the electrostatic system is fully controllable at the nominal operating point. . SOLUTION The controllability matrix (Equation 4. 9. guaranteed.4. While observability is. À700 V (lower curve)].121) is given by Q ¼ [B AB] ¼ 0:9798 0 0:9798 0:3315 11:1026 0:3315 0 11:1026 ! Since the rank of the controllability matrix is two.Internal Process Controls 0 –50 –100 –150 Vl (volts) 483 –200 –250 –300 –350 –400 –450 2 4 6 8 10 X (ergs/cm2) 12 14 16 FIGURE 9. We now check these conditions for the LTI system of Equation 9. À600 V. the same conclusion can be reached by ﬁnding the rank of the Jacobian matrix (B).

P).2. time-invariant state-space model for the error and we can control the decay rate of the components of vector e(k) by properly selecting the gain matrix K to place the eigenvalues (poles) of the matrix (I À BK) at positions inside the unit circle in the complex z-plane to meet the required stability speciﬁcations of the closed-loop system (Section 5.3 Using the gain matrix of Example 9. The gain matrix K is computed to be ! 0:8165 0 K¼ À0:0244 0:0630 This gain matrix will give rise to stable performance. via Equation 9. command K ¼ place (A.2 For the open-loop electrostatic system shown in Example 9. we end up having an autonomous linear.3). say xd.2 and 0. A robust poleplacement algorithm for a MIMO electrostatic system is called ‘‘place. simulate the transient performance of the closed-loop linear .3). Example 9. SOLUTION Using A ¼ I.3] in MATLAB. Example 9. ﬁnd the gain matrix K to place the closed-loop poles within the unit circle on the real axis at 0.2 and the electrostatic model described in Chapter 10. Other aspects of closed-loop performance can be also used in designing K (see Section 5. It minimizes the sensitivity of closed-loop poles to uncertainties in the A and B matrices (see Example 5.1. Reference [13] gives an additional procedure for assigning poles to a closed-loop MIMO system. which will be discussed later in this chapter. we can deﬁne the error signal as e(k) ¼ xd À x(k) (9:25) Then.8 for designing level 2 controller gain matrix is also applicable to the level 1 controller gain matrix.21 we have the following equations for the closed-loop system e(k þ 1) ¼ e(k) À Bv(k) v(k) ¼ þKe(k) or equivalently e(k þ 1) ¼ (I À BK)e(k) (9:27) (9:26) Therefore.5). B ¼ 0:9798 0 0:3315 11:1026 ! and P ¼ [0.484 Control of Color Imaging Systems: Analysis and Design In the case of tracking of a reference value.3. which in our case are the uncertainties in the elements of the Jacobian matrix.’’ which can be found in the MATLAB Control System Toolbox [12]. which uses an extra degrees of freedom to ﬁnd a robust solution for gain matrix K. A simple approach described in Section 9.2 0. B.

can maintain the exposure level to the desired value inspite of the dark decay as seen in the next example. Clearly.Internal Process Controls –550 –555 –560 –560 –565 –570 485 Vh (volts) –565 –570 –575 –580 –585 –590 0 2 4 6 8 10 12 14 Vg (volts) –575 –580 –585 –590 –595 –600 0 2 4 6 8 10 12 14 (a) Pitch number (b) Pitch number FIGURE 9.8 and 9. .8 6. The length of the time that the charge is retained on a photoconductor is determined by the decay rate in the dark. if designed to be stable. Charge and exposure do not take place instantaneously. –86 –88 –90 8 7. the transient performance is similar to the performance expected by a system with closed-loop poles at 0.9 (a) Convergence plot of Vl with respect to pitch number. electrostatic control system when the desired unexposed and exposed voltages on the photoconductor are given by xd ¼ [À550 À100]T volts.2 and 0.8 (a) Convergence plot of Vh with respect to pitch number.6 6.6 V1 (volts) –92 –94 –96 –98 X (ergs/cm2) 7. (b) Plot of actuator X(Ul) with respect to pitch number. SOLUTION Figures 9.4 7. (b) Plot of actuator Vg(Ug) with respect to pitch number. that is. except for a small overshoot in Vl. with closed-loop eigenvalues within the unit circle.2 7 6.4 6. Automatic control systems.9 show the convergence of the unexposed and exposed voltages as the actuators (grid voltage and laser intensity) are changed every photoconductor pitch.8 7.2 6 –100 –102 0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 (a) Pitch number (b) Pitch number FIGURE 9. Sometimes the period between charge and exposure can lead to signiﬁcant loss of charge.3.

5 7 6.e. whereas the starting value is different because the photoconductor was charged to a lower potential due to dark decay.10).5 6 5.4 Let the charge on the photoconductor at the exposure station be decayed by 30 V (i.2a.2b which is being used extensively in this study. but the motivation of this exercise is to prove that pole-placement algorithms for both cases have an identical form. which in our case reduces to the simple condition that the Jacobian matrix (at a nominal operating point) has full rank. The system is assured to be LTI. In the next section. The result is well known.. at this point. Vl has the desired ﬁnal value. Interestingly. Vh at the exposure station is equal to À550 þ 30 V). Of course.5 STATE SPACE TO TRANSFER FUNCTION CONVERSIONS In this section. 9.3. X has the same starting value but a different ﬁnal value when compared to Example 9. that all the pole-placement algorithms are designed for continuous-time linear systems of the form shown in Equation 9. we demonstrate how to transform a system from state space to transfer function form. Note that it is known from linear system theory that for arbitrary pole placement to be possible.486 –70 –75 –80 –85 Control of Color Imaging Systems: Analysis and Design 8 7. SOLUTION Vh and Vg plots do not change. Simulate the transient performance of the controller shown in Example 9. It is interesting to note. we know from linear system theory .5 0 2 4 6 8 10 12 14 5 0 2 4 6 8 10 12 14 Vl –90 –95 –100 –105 Print number X Print number FIGURE 9. Example 9. a necessary and sufﬁcient condition is controllability. we show that they are equivalent to pole-placement techniques for the discrete model Equation 9.10 Convergence plot of V1 and actuator X(Ul) as a funtction of print number for a one pitch photoconductor with dark decay of 30 V at the exposure station. respectively. the electrostatic control system is automatically able to maintain the photoconductor voltages to the desired target (Figure 9.9. particularly for the continuous-time case. This is done for both continuous and discrete-time systems using the Laplace transform (s-domain) and z-transform.

an LTI continuous-time system can be written as _ x ¼ Ax þ Bu y ¼ Cx þ Du or in the discrete-time x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) þ Du(k) Taking the Laplace transform of both sides of Equation 9. and where the inverse (zI À A)À1 exists for all values of z except at the eigenvalues of matrix A. In recapitutation. u(k). (9:31) (9:30) (9:29) (9:28) . y(k). respectively. Therefore. and U(s). we get zX(z) À x(0) ¼ AX(z) þ BU(z) Y(z) ¼ CX(z) þ DU(z) ) X(z) ¼ (zI À A)À1 x(0) þ (zI À A)À1 BU(z) Y(z) ¼ CX(z) þ DU(z) ¼ C(zI À A)À1 x(0) þ (C(zI À A)À1 B þ D)U(z) where X(z). and U(z) denote the one-sided z-transforms of the signals x(k). Y(s).Internal Process Controls 487 that discussion on arbitrary pole placement is only possible if the controllability condition is satisﬁed.28. and where the inverse (sI À A)À1 exists for all real values of s except at the eigenvalues of matrix A. only s is replaced by z. a poleplacement algorithm to design state feedback for a continuous-time system can also be used for the discrete case. This shows that the two (matrix) transfer functions have exactly the same form. we get sX(s) À x(0) ¼ AX(s) þ BU(s) Y(s) ¼ CX(s) þ DU(s) ) X(s) ¼ (sI À A)À1 X(0) þ (sI À A)À1 BU(s) Y(s) ¼ CX(s) þ DU(s) ¼ C(sI À A)À1 x(0) þ (C(sI À A)À1 B þ D)U(s) where X(s). Y(z). y(t). denote the Laplace transforms of the signals x(t). u(t). Similarly.29. applying one-sided z-transform on Equation 9. since the transfer function has the exact same form in both cases. respectively.

. the development voltage. A copy produced by an overcharged photoconductor will have a gray or dark background instead of the white background of the paper. Vdev.488 Control of Color Imaging Systems: Analysis and Design Hence. that is. to achieve stability of the closed-loop system. for a discharged area development. yellow. Vclean which affects the background and dot growth of the image on the photoreceptor. where K is a constant gain matrix chosen to satisfy the speciﬁcations (i.30) will be relatively weak and the resulting deposition of toner material will be weakened.29 and 10. then the electrostatic latent image obtained upon exposure of the image (e. Of course.e. again s is replaced with z. toner is developed on an exposed dot to create a solid cyan. Therefore. shown in Equations 10. As a result. we can use the same pole-placement algorithms developed for continuous-time LTI systems. If. or black dot for a discharge area development.g. however. if the photoconductor is excessively overcharged. there will be too much developer material that is deposited on the photoconductor. A binary printer uses two levels with a dot ‘‘on’’ or ‘‘off’’. the poles have to be placed in the left-hand side of complex s-plane for continuous-time systems and inside the unit circle of the complex z-plane for discrete-time systems. the copy produced by an underexposed photoconductor would look faded since Vdev will not be sufﬁcient for good development. In addition to the background effect. If the photoconductor is not well controlled for both fully exposed (V1—the voltage on the discharged region) and unexposed levels (Vh—the voltage on the charged area). the photoconductor is overcharged. it can be permanently damaged.. 9. areas intended to be gray will be black and the tone reproduction will be poor. Moreover. magenta. In some machines multiple exposure levels are obtained by performing intensity modulation during exposure. taking u(t) ¼ þKx(t). In a digital printing machine such as an electrophotographic engine. This results in a higher cleaning voltage.6 LEVEL 2 DEVELOPABILITY CONTROLLER Control of the ﬁeld strength and the uniformity of the charge on the photoconductor is very important because high-quality prints are produced when a uniform charge with the desired magnitude is maintained on the photoconductor. so that the closed-loop system has the desired eigenvalues) we get X(s) ¼ (sI À A À BK)À1 x(0) Y(s) ¼ (C þ DK)(sI À A À BK)À1 x(0) or in the discrete case: X(z) ¼ (zI À A À BK)À1 x(0) Y(z) ¼ (C þ DK)(zI À A À BK)À1 x(0) (9:33) (9:32) that is. each tone of a contone image is produced by a certain spatial combination of the available tone levels produced by the halftone screening process.. These levels are used to reproduce the contone image.

The TC changes during engine operation due to a varying amount of toner depletion (or addition) caused by image development. iGen3) use ﬁxed area coverages. and development bias at some periodic intervals. During control. for time-sequential sampling of tone curves. a direct control of Q=M ratio cannot be implemented easily without increased hardware cost. one tone level (area coverage) is proposed during a periodic actuation instance. Another way of reducing the variability in development is to measure the DMA with sensors and adjust the tone levels using area modulation as in the level 3 control loop described later in Section 9. A common technique for measuring DMA is to artiﬁcially create test patches of reasonable size (e. the DMA measurement vector from the sensor is ﬁltered using an appropriate algorithm to remove noise components and obtain a true . Some printers. the quality of the developed image on the photoconductor depends on the system’s ability to reproduce the tones in the presence of exposure and development uncertainties. This toner mass is a function of numerous parameters of which many are ﬁxed by design.Internal Process Controls 489 Although halftoning process is well deﬁned and repeatable.16]. and toner agitation mechanism in the developer. Other more feasible implementations (e. For example.2 on a photoconductor belt) in the IDZ at predetermined area coverages that act as a surrogate measurement for capturing effects on the customer image area [9. which develop the image on four individual drums. The developed test patches are read by the reﬂected signal from the patch area [14]. Generally. 1 in. The developed mass of toner per unit area (DMA) is normally used to quantify the macro level toner reproducibility on the photoconductor. The number of test patches used depends on how many tone levels in the tone curve have to be sampled to generate the control actuation and the approach used to implement the control function. TC is maintained to some constant level using a sensor in a separate multilevel asynchronous control loop described in Section 9. and development bias [1.9 and=or by adjusting photoconductor charge.11. Spatial effects are captured as an integrated signal in the sensor output and they contribute toward noise in the measurement.g. one at low area coverage (0%–20%) and one at mid-tone (around 50%). use a two patch sampling strategy. in xerographic engines. Parameters that contribute to the variability of DMA within print runs can lead to print quality problems and increase the number of misprints. TC. such as the DC2060 and DC6060.10]. and developed by the laser system as necessary. which is considered adequate to sample the critical regions of the tone curve and affect the entire curve through feedback to process actuators such as charge. relative humidity (RH) of the air in the development substation. exposure intensity. one at high area coverage (90%–100%). exposed. The charge-to-mass ratio of the toner is dependent on the triboelectric properties of the toner and carrier.. exposure. Although development can be affected directly by varying the Q=M ratio. A three patch sampling is required at reasonably high print frequency to maintain the stability of mid-tones in image-on-image engines.. The most important parameters in a two component developer system are (1) toner concentration (TC)—the ratio of the amount of toner to the amount of carrier available in the development system and (2) the toner charge per unit mass (Q=M ratio). These test patch areas are charged. [15]. The patch area coverage is varied over time using a time-sequential sampling strategy.g. in Ref.

and Ub (k) ¼ Vbias (k) are captured by the actuator vector U(k) ¼ [Uh (k) Ul (k) Ub (k)]T . Let Dl(k). signal. k is a parameter used to describe the iteration of the level 2 control loop. the set points for the level 1 loop {Vh and Vl} act as two actuators. The choice of actuators depends on the control hierarchy and the ease of implementation.11 MIMO level 2 developability controller.11.490 Control of Color Imaging Systems: Analysis and Design Filtered sensor output (Low. The developer bias. and Dh(k) represent the three different DMA measurements (output of the ﬁlter) at the sample time deﬁned by the parameter k. U0 Actuations to level 1 and development system FIGURE 9. 10–30 prints are made for every level 2 iteration. Elements in the matrix represent the ﬁrst-order derivatives of the DMA of the three patches with respect to the three actuators. . xd ¼ [DT DT DT ]T . For a three patch sampling system. The actuator signals Uh (k) ¼ T Vh (k). Dm(k). K v(k) Integrator u(k) U(k) Xerographic system with level 1 + Targets (Low. and high DMAs). since the electrostatic control loop is fully settled before closing the outer level 2 control loop. and high DMAs). Ul (k) ¼ VlT (k). This parameter is different from the iteration number of the level 1 loop. the simplest controller would be a three-input three-output state feedback controller whose gains are determined from the Jacobian matrix. mid. The design of a MIMO controller is described in Figure 9. Numerous additional architectures use combinations of process parameters as actuators [16–21] in order to optimize for cost and performance. The ﬁltered DMA vector is compared to the desired DMA target vector and the error vector is processed by the feedback controller to generate the actuator values. Let the corresponding target values of the toner mass be represented by the vector. Vbias. The target vector does not depend on k since there is no higher l m h level loop used to update the target vector. mid. In a time hierarchical control architecture. Normally. xd + Nominal actuator values. in which the sensed signals are coupled to all of the actuators. is used as the third actuator. One iteration pass for a level 2 loop comprises of a sensing-processing-actuation update cycle. Several iterations of the inner loop are required to reach steady performance before the outer loop is closed.11. Let U0 ¼ [Uho Ulo Ubo ]T be the corresponding static nominal actuator vector that does not depend on the sample time. These DMA measurements are basically identiﬁed as three different points on the TRC. The hierarchical structure contains nested loops: an inner loop (level 1) operating within the body of an outer loop (level 2). x(k) – e(k) Gain matrix. The linear state feedback controller will have the architecture shown in Figure 9. which affects how much toner is transferred to the photoconductor.

Internal Process Controls

491

The actuator updates are obtained from U(k) ¼ U0 þ u(k) u(k þ 1) ¼ u(k) þ v(k) This closed-loop system equation is very similar to Equation 9.26. (9:34)

9.6.1 JACOBIAN MATRIX

FOR

DEVELOPABILITY CONTROL

Development of the Jacobian matrix for level 2 control loop is very challenging because there are three actuators being coupled to the sensor signals as apposed to two actuators in the case of electrostatic control. We can think of three methods for extracting the Jacobian matrix. They are (1) an analytical method, (2) design of experiments (DoE) methods [22,23], and (3) direct numerical methods. In the analytical method, at ﬁrst the Jacobian matrix of a development system is expressed in terms of analytical expressions for the separation of interest using a charging, exposure, and development model. The Jacobian with respect to the actuators of the process models can be calculated using the symbolic toolbox in MATLAB. Apart from the Jacobian matrix, this approach gives invaluable information about the robustness of the model with respect to key process parameters of the print engine. This information can be used to satisfy several other speciﬁcations and improve the system performance. The direct experimental method involves running carefully designed experiments to extract the nonlinear surfaces in the DMA outputs with respect to actuator values. The DoE method allows systematic ways to select actuator values in the vector, U0 ¼ ½Uho Ulo Ubo 0 , between their operating limits, so that they are orthogonal and yet minimize the number of experiments required to determine the Jacobian matrix. We will describe the numerical method in detail below. The numerical method will be useful to understand the available operating space for the DMA vector in a static machine when actuators are varied within their limits. Figure 9.12 shows a three-dimensional (3-D) space covered by the 163 actuator values for combinations sampled uniformly for a tandem printer example. Each point on this plot represents the actuator vector, U. In DoE language, this type of design matrix is called a ‘‘full factorial design.’’ In Operations Research, this is known as an exhaustive search. Three patches with low, mid, and high area coverages are developed under different actuator values for U spread over their limits. They are measured with the optical DMA sensor, if actual printer is used (or calculated when a virtual printer model is used). Figure 9.13 shows the volumetric space covered by the DMA vector for the actuator values of Figure 9.12 shown for four different orientations of the DMA axes. Figure 9.14 shows the full development TRC for the same actuator values. Clearly, the available space for controlling the DMA using the hierarchical level 2 controls is quite limited. The high DMA values show a larger change, meaning higher sensitivity to actuator changes. However, careful selection of actuators can further increase the available space, which will be covered brieﬂy in Example 9.7 and Section 9.14.

492

–150

Control of Color Imaging Systems: Analysis and Design

Vbias (volts)

–200

–250

–300 –50 –60 Vlow (volts) –70 –80 –90 –100 –700 –650 –600 Vhigh (volts) –550 –500

FIGURE 9.12

Actuator space used for identifying the numerical DMA model for a static printer.

0.5 0.45 0.5 DMA high DMA high 0.3 DM 0.2 Am id 0.1 0.02 0.06 0.04 low DMA 0.08 0.4 0.3 0.2 0.1 0.4 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.4 0.2 DMA 0 mid

0 0

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 DMA low

0.5 0.45 DMA high 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0

A DM

0.5 0.45 0.4 DMA high 0.2 0.4 0 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 DMA low 0.35 0.3 0.25 0.2 0.15 0.1 0.4

0.3 0.2 0.1 DMA mid

0 0.04 0.02 0 0.08 0.06 DMA low

FIGURE 9.13 DMA vector shown as points along the DMA low, mid, and high axes (DMA values are shown in mg=cm2 units).

d mi

**Internal Process Controls
**

0.45 0.4 0.35 DMA (mg/cm2) 0.3 0.25 0.2 0.15 0.1 0.05 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Area coverage 0.7 0.8 0.9 1

493

FIGURE 9.14 Development TRCs for the actuator space shown in Figure 9.12.

The input–output data is used for determining the Jacobian matrix. A numerical 3-D model is constructed through interpolation. Nearest neighbor, linear, spline, or cubic interpolation methods can be used, depending on the input–output data. Key steps in determining the Jacobian matrix at the nominal operating points are 1. Set the actuator vector, U0 ¼ [Uho Ulo Ubo ]0 to the desired nominal value. 2. Vary Uh to some value about the nominal value of Uho (increase by that value and decrease by that value) while keeping other actuator values at their nominal operating points. Record the difference DMA vector for these actuator settings obtained using the numerical model. This step gives the ﬁrst column vector of the Jacobian matrix. 3. Vary Ul to some value about the nominal value of Ulo while keeping other actuator values at their nominal operating points. Record the difference DMA vector for these actuator settings obtained using the numerical model. This step gives the second column vector of the Jacobian matrix. 4. Vary Ub to some value about the nominal value of Ubo while keeping other actuator values at their nominal operating point. Record the difference DMA vector for these actuator settings obtained using the numerical model. This step gives the last column vector of the Jacobian matrix. Example 9.5

Test the following developability control system for complete controllability. The system is given at the nominal operating point {Uho ¼ À600 V, Ulo ¼ À100 V, Ubo ¼ À300 V}.

494 2 1 0

Control of Color Imaging Systems: Analysis and Design 0 3 2 0:0282 0:5024 À0:2633 3

6 x(k þ 1) ¼ 4 0 1 0 0 2 1 0 6 y(k) ¼ 4 0 1 0 0

7 7 6 0 5x(k) þ 10À3 4 0:1413 1:6703 À0:9165 5v(k) 0 1:2781 À1:2854 1 3 0 7 0 5x(k) 1

SOLUTION

The controllability matrix is given by

Â Ã Qo ¼ B AB A2 B 2 3 0:0282 0:5024 À0:2633 0:0282 0:5024 À0:2633 0:0282 0:5024 À0:2633 6 7 ¼ 10À3 6 0:1413 1:6703 À0:9165 0:1413 1:6703 À0:9165 0:1413 1:6703 À0:9165 7 4 5 0 1:2781 À1:2854 0 1:2781 À1:2854 0 1:2781 À1:2854

The rank of the controllability matrix is equal to 3. That is, the developability system with level 1 loop is fully controllable at the given nominal operating point. In other words, there are three independent factors involved in this process (due to rank of the Jacobian matrix), and all of them can be controlled using three actuators in vector v(k). That means, a gain matrix K of size 3 Â 3 can be designed to affect all three DMAs using an arbitrary pole-placement strategy.

**9.7 STEADY-STATE ERROR
**

An important part of the level 1 and 2 controllers is that the closed system has the ability to give zero (or minimum) steady-state error, theoretically, in a single measurement-actuation update cycle. The steady-state error is deﬁned as the error between the measured (ﬁltered) values and the desired target values. Due to the use of an integrator in the loop we show mathematically below how the steady-state error is driven to zero with multiple iterations. Consider the open-loop state Equation 9.21, error Equation 9.25, and the feedback Equation 9.26. The state-space equations are very similar in structure for both level 1 and 2 controllers, but they have different states, Jacobian, dimensionality, output, etc. Substituting Equation 9.26 in Equation 9.21 we get the closed-loop state equation as follows: x(k þ 1) ¼ x(k) þ B[Ke(k)] Substituting the error vector, Equation 9.25, in Equation 9.35, we get x(k þ 1) ¼ (I À BK)x(k) þ BKxd Taking the z-transform of Equation 9.36, we have zX(z) À zx0 ¼ (I À BK)X(z) þ BKXd (z) (9:37) (9:36) (9:35)

Internal Process Controls

495

where x0 is shown in Equation 9.13, and is the value of the state vector when the actuators are set equal to the nominal values, u0. Xd(z) is the z-transform of the desired target vector. Now, by rearranging Equation 9.37, we can write the expression for X(z) as follows: X(z) ¼ [zI À (I À BK)]À1 BKXd (z) þ z[zI À (I À BK)]À1 x0 (9:38)

where matrix I is the identity matrix. Now the error vector can be written in z-domain as follows: E(z) ¼ Xd (z) À X(z) ¼ [I À {zI À (I À BK)}À1 BK]Xd (z) À z[zI À (I À BK)]À1 x0 (9:39)

The reference vector containing desired target values can be written in z-domain as follows: Xd (z) ¼ z Xd zÀ1 (9:40)

Using ﬁnal value property, Equation 3.70, we can show that the steady-state error is driven to zero after multiple actuations. ess (k)jclosed ¼ lim (z À 1)E(z)

z!1

(9:41)

Substituting Equations 9.39 and 9.40 into Equation 9.41 and applying the limit as z ! 1, the steady-state error will make its way to zero. This clearly shows that when the system is stable (very important condition!) the output will settle to the desired target with minimum error equal to or close to zero. This is true regardless of the Jacobian=gain matrix and even the initial state, x0 as long as the feedback loop is stable. In contrast, for the open-loop case, the following analysis shows that the steadystate error is not independent of the initial state, x0, because the actual states will drift with time (i.e., actual x(k) will be different from x0). Applying z-transform on the open-loop state equation, we can write the state vector in z-domain as follows: X(z) ¼ B z V(z) þ x0 z zÀ1 (9:42)

For the open-loop case the vector, V(z) ¼ 0. Now the expression for the error vector between the desired inputs to the outputs without feedback can be written as follows: E(z) ¼ Xd (z) À X(z) z B ¼ ðXd À x0 Þ À V(z) zÀ1 z

(9:43)

496

Control of Color Imaging Systems: Analysis and Design

Using ﬁnal value property, we have ess (k)jopen ¼ Xd À x0 (9:44)

We ﬁnd that the steady-state error is a function of the state vector, x0 and is not equal to zero unless u0 creates the same value as the desired vector, Xd, which is hardly the case since there is always drift=media change happening during system operation.

**9.8 DESIGN OF THE GAIN MATRIX
**

The design of the gain matrix is important for two reasons: (1) for controlling the convergence rate (or number of iterations) for the states to a steady-state value and (2) for staying within the stability bounds so that even when the toner mass deviates due to various disturbances (e.g., toner usage change, media change, developer aging, dark decay, etc.) in the print engine, the closed-loop system should move the states to the desired steady-state value. In this section, the design of gain matrix using a limited pole-placement design technique is discussed. There is no simple generalized procedure for pole-placement design of MIMO systems. Since the degree of freedom available to choose the right kind of gains is sufﬁciently large, one can assign performance measures to pick the gains (e.g., to improve the output states in the presence of noise), to assign eigenvectors that would respond according to the natural modes of the system. Such techniques are more elaborate and difﬁcult. Optimal controls and various robust pole assignment techniques are readily available in the literature [13,24] and discussed in Chapter 5 to perform those calculations using the state variable formulation described above. For a simple pole-placement design for level 2 system, let us choose the gain matrix in such a way that the eigenvalues of the closed-loop system matrix, A-BK, shown in Equation 9.36 has the desired closed-loop poles. To select desired poles for a discrete system, see Refs. [13,24,25]. The location of the closed-loop poles dictates the required number of iterations for driving the error states toward zero. Since A ¼ I for the level 1 and 2 system, the gain matrix can be easily written in terms of closedloop poles as K ¼ BÀ1 s (9:45)

where s is a 3 Â 3 matrix determined by the closed-loop poles. For a stable system, all the poles should be assigned values between 0 and 1. For a three-input threeoutput control system, s is written in terms of three poles (p1, p2, p3) as follows: 1 À p1 s¼4 0 0 2 0 1 À p2 0 3 0 0 5 1 À p3 (9:46)

Theoretically, for the system with A ¼ I, we can show that when all three closed-loop poles are chosen real and equal, the gain matrix will result in a more robust closedloop performance.

Internal Process Controls

497

Since the Jacobian matrix is a function of the actuator for the level 2 system, the gain matrix can be computed for different regions of the actuator operating space. That is, the gain matrix will be actuator dependent, which can be written as À Á À Á Kj Uj ¼ BÀ1 Uj s j (9:47)

where sufﬁx j represents the discrete nature of the actuator vector. It also means that the operating actuator space is quantized to j ¼ 1, 2, 3, . . . , J regions. This type of pole assignment will work when the Jacobian matrix is invertible. If faster convergence is required, then it is normal practice to assign the poles (p1, p2, p3) close to 0. If the poles are assigned equal to zero, then s ¼ I, and the performance will be theoretically ‘‘dead beat.’’ That means, due to the multivariable nature of the system, the theoretical minimum for convergence to zero steady-state error will be ‘‘one iteration’’ (i.e., one measurement-actuation update). If the poles are assigned closer to unity (but not exactly unity, because unit pole values will lead to marginal stability, meaning no convergence), then number of iterations will be greater than three (about 7 to 8). As a result, the performance of the closed-loop system in the presence of system or sensor noise can be improved. Example 9.6

Let the desired closed-loop poles for level 2 controller be located at p1, p2, p3. (i) Using Equation 9.45 for the Jacobian matrix of Example 9.5 ﬁnd the gain matrix that will provide equal poles p1 ¼ p2 ¼ p3 ¼ 0.3. Use the MATLAB pole-placement algorithm from Ref. [12]. Find the gain matrix using both methods for poles at p1 ¼ 0.3; p2 ¼ 0.3; p3 ¼ 0.7. What do you see? Calculate the step responses for case (i). Compare the results.

(ii)

SOLUTION

Case (i): Gain matrix from Equation 9.45 for pole values p1 ¼ p2 ¼ p3 ¼ 0.3 is given by 3 À4:223 1:3384 À0:0891 K ¼ 104 4 0:7863 À0:1570 À0:0492 5 0:7818 À0:1561 À0:1033 The MATLAB pole-placement algorithm gives the same gain values for equal poles. Equation 9.45 and the MATLAB pole-placement algorithm give the same gain matrix for pole values p1 ¼ 0.3; p2 ¼ 0.3; p3 ¼ 0.7. 2 3 À4:223 1:3384 À0:0382 44 K ¼ 10 0:7863 À0:1570 À0:0211 5 0:7818 À0:1561 À0:0433 Since the system matrix, A, is 3 Â 3 and is equal to identity, Equation 9.45 gives a robust gain matrix with all the beneﬁts of the pole-placement algorithm, place(). 2

498

Control of Color Imaging Systems: Analysis and Design

This is true even for the level 1 controller gain matrix. Also, notice the changes between two gain matrices. They are in the last column of the gain matrix which is due to change in p3. The third pole p3 controls how the DMA transients are affected due to the actuator, Vbias. Case (ii): Step responses for the DMA states were created by changing the DMA vector from [0.0384 0.1301 0.3398] to [0.0384 0.1301 0.3738] for two different gain matrices shown in case (i). Level 2 controller has sensing-processingactuation updates at every 30 pitch, whereas the inner level 1 loop is updated every pitch. DMA vector shown in Figure 9.15a converges to the desired steadystate value after few updates. The actuator response is shown in Figure 9.15b. Figure 9.15c shows the response of the inner loop when the level 2 control loop is in its last iteration pass.

Example 9.7

The cost of the electrostatic sensing system can be reduced by removing the charge sensor (ESV) and associated level 1 control loop. It is now required to design a three-input three-output developability control loop using three process actuators, the grid voltage, Ug(k) ¼ Vg(k), the exposure intensity, Ul(k) ¼ X(k), the development bias, Ub(k) ¼ Vbias(k), and measurements from the DMA sensor. Let Dl(k), Dm(k), and Dh(k) represent the three different DMA measurements shown in the state vector, x(k), measured every pitch, indicated by the parameter, k. The linear state variable description of the control system is characterized by the Jacobian matrix at the nominal operating point {Ugo ¼ À600 V, Ulo ¼ 8 erg=cm2, Ubo ¼ À300 V} as shown below: 1 x(k þ 1) ¼ 4 0 0 2 1 6 y(k) ¼ 4 0 2 2 0 1 0 0 1 3 2 0 0:2789 À3 4 0 5x(k) þ 10 0:9761 1 0:4225 3 0 7 0 5x(k) 1 3 À0:2777 8:7889 À1:0703 29:7035 5v(k) À1:2727 16:8863

0 0

3 2 3 Dl (k) DUg (k) where x(k) ¼ 4 Dm (k) 5; v(k) ¼ 4 DUl (k) 5; y(k) ¼ x(k). The controller uses a gain Dh (k) DUb (k) matrix and the integrator modeled by following equation: e(k þ 1) ¼ e(k) À Bv(k) v(k) ¼ þKe(k) where e(k) ¼ xd À x(k), and xd is the desired state vector. i. ii. Find the gain matrix for placing the poles at location [0.3, 0.3, 0.3] using Equation 9.45. Show the time evolution of the states for a step response as a function of pitch number using recursive solution of equation (Section 4.8.3), the measurement-actuation updates are executed at every pitch.

0.0398 0.137 0.370 0.365 0.360 0.355 0.350 0.345 0.340 5 10 15 35 40 45 50 0.335 0 5 10 15 35 40 45 50 0.136 0.135 0.134 0.133 0.132 0.131 0.130 45 20 25 30 30x Pitch # 50 0.129 0 DMA high

0.138

0.375

0.0396

0.0394

0.0392 DMA mid

DMA low

0.0390

0.0388

Internal Process Controls

0.0386

0.0384

0

5

10

15

35

40

(a)

20 25 30 30x Pitch #

20 25 30 30x Pitch #

–580 –100 –105 –110 –115 –120 –125 0 5 –310 –320 Vbias (volts) –330 –340 –350 10 15 20 25 30 35 40 45 50 30x Pitch # –360 0 –300

–600

–620

–640 Vlow (volts)

–660

Vhigh (volts)

–680

–700

–720

–740 0

5

(b)

10 15 20 25 30 35 40 45 50 30x Pitch #

5

10 15 20 25 30 35 40 45 50 30x Pitch #

499

FIGURE 9.15 (a) Responses of the DMA vector. Samples are shown at 30 pitch interval. (b) Responses of the actuator vector. Samples are shown at 30 pitch interval. Solid curve represents P1 ¼ P2 ¼ P3 ¼ 0:3 and dashed curve represents P1 ¼ P2 ¼ 0:3 and P3 ¼ 0:7. All DMA values are shown in mg=cm2 units. (continued )

500

–580 –70 –80 –90 –100 –110 –120 10 15 20 Pitch # 25 30 –130 0 5 10 15 20 Pitch # 25 30 Vl (volts)

–60

–600

–620

–640

Vh (volts)

–660

–680

–700

–720

–740 0

5

(c)

Control of Color Imaging Systems: Analysis and Design

FIGURE 9.15 (continued)

(c) Responses of the level 1 state vector (Vh, Vl) with respect to every pitch when level 2 is in its last iteration pass.

Internal Process Controls iii. iv.

501

Compare the results of step ii by running the control simulation with the charge and development models from Chapter 10. Are there other reasons why this type of control approach is more suitable than level 1 and 2 architecture? Justify your answers.

SOLUTION

Case (i): Gain matrix from Equation 9.43 for pole values p1 ¼ p2 ¼ p3 ¼ 0.3 is given by 2 3 À3:9405 1:2977 À0:2317 44 0:7855 À0:1988 À0:0592 5 K ¼ 10 0:1578 À0:0475 0:0055 Gain matrix from Equation 9.43 for pole values p1 ¼ 0.3; p2 ¼ 0.3; p3 ¼ 0.7 is given by 2 3 À3:9405 1:2977 À0:0993 44 0:7855 À0:1988 À0:0254 5 K ¼ 10 0:1578 À0:0475 0:0023 Case (ii): Step responses for the DMA states were created by changing the DMA vector from [0.0384 0.1301 0.3398] to [0.0384 0.1301 0.3738]. The DMA vector shown in Figure 9.16a converges to the desired steady-state value for the two different gain matrices. The actuator response is shown in Figure 9.16b. Solutions for cases (iii) and (iv) are left to the reader.

**9.9 LEVEL 3 CONTROL LOOPS
**

At level 3, the TRCs of the individual separation are maintained in such a way that they correspond to a reference tone curve on the photoconductor or on the paper with the fused toner. The tone curve on the photoconductor (called ‘‘developed reproduction curve’’) is a function that maps digital contone values of a color separated image to DMA values on the photoconductor. Similarly, the tone curve on the paper is a function that maps a color separated image to an optical density on the paper after fusing. For black and white printers, what matters is the tone curve on the paper. To achieve good reproduction, it is important that the desired curve be linear. Often, users request different tone curves depending on the overall appearance of the image. For four-color digital reproduction, the complex interaction of each of the tone curves on the paper makes it difﬁcult to specify the exact shape of the desired tone curves. A linear reference curve is more widely used in the industry. By design, the DMA at 100% area coverage is generally limited to a set point and density control loops try to maintain solids at those points. As a result, the available control space is externally limited for level 2 process controls, which involves achieving the desired linear tone response on the photoconductor or paper. The controllable space is illustrated conceptually in Figure 9.17. In this ﬁgure, level 2 targets are in the controllable space that is outside the space where the desired reference TRC curve, which is linear in this illustration. Thus, the desired tone curve is uncontrollable by levels 1 and 2 actuators. For such situations, a preferred way to achieve the desired tone response is through

502

0.0405 0.135 0.134 0.365 0.133 0.36 0.355 0.35 0.345 0.13 0.34 5 0.335 0 5 0.129 0 10 15 20 25 30 35 40 45 50 Pitch # 0.132 0.131 0.37

0.375

0.04

0.0395

DMA low

0.0385

0.038 0

5

(a)

10 15 20 25 30 35 40 45 50 Pitch #

DMA mid

DMA high

0.039

10 15 20 25 30 35 40 45 50 Pitch #

–600 –300 –305 –310 –315 –320 –325 8.5 8 7.5 7 6.5 5 6 0

9

–620

–640

–660 –330

Vgrid (volts)

Vbias (volts)

–700 –340 –345 –350 0

–335

–720

–740 0

5

X (ergs/cm2)

–680

5

(b)

10 15 20 25 30 35 40 45 50 Pitch #

10 15 20 25 30 35 40 45 50 Pitch #

10 15 20 25 30 35 40 45 50 Pitch #

Control of Color Imaging Systems: Analysis and Design

FIGURE 9.16

(a) Responses of the DMA vector (all DMA values are shown in mg=cm2 units). (b) Responses of the actuator vector.

Internal Process Controls

503

Controllable space for level 2 Actual TRC 100% % TRC (output) 75% Ideal TRC curve 50%

25%

25%

50% 75% % Area coverage (input)

100%

FIGURE 9.17 Schematic diagram representing the TRC with controllable space using process actuators (charge and development bias).

Actual TRC 100% % TRC (output) 75%

50%

25%

Compensation or TRC inverse 25% 50% 75% % Area coverage (input) 100%

FIGURE 9.18 Schematic diagram representing the inverse TRC obtained to achieve the desired reference TRC.

level 3 controls, since these higher level controls can provide extra degrees of freedom that lower level (charge and developability) controls cannot. The level 3 control system generates the inverse curve (Figure 9.18) for each separation [26]. This curve can be obtained by inverting the measured TRC over a reference TRC. In another scenario presented in Figure 9.19, it appears as though there is sufﬁcient actuator space to obtain the desired tone curve. However, we are still not able to reach full inversion with three point level 2 controls because of insufﬁcient sampling; sampling of only three (low, mid, and high area coverages) is sufﬁcient for inversion. A time-sequential sampling proposed in Ref. [15] may provide reliable control for such systems for a given area coverage (or gray level); during each

19 Left: Schematic diagram representing the TRC with controllable space covering the reference TRC using process actuators (charge and development bias).504 Controllable space for level 2 Actual TRC 100% x 75% 50% 25% x Actual TRC After level 2 controls 75% 100% 100% 75% % TRC (output) 25% x 50% % TRC (output) 50% Compensation or TRC inverse 25% % Area coverage (input) 25% 50% 75% 100% % Area coverage (input) Ideal TRC and level 2 targets Control of Color Imaging Systems: Analysis and Design FIGURE 9. . Right: Schematic diagram representing the inverse TRC achievable using contone actuators.

Also. but this results in too many patches and may be unnecessary when the TRC has less structure. depending on the measurement resources supplied by the print engine. then the inverse TRC is obtained using the ‘‘static inverse’’ techniques illustrated graphically in Figures 9. The normalized DMA or DE from paper (referred to as the output tone value) can have values lower or higher than the reference curve as illustrated in Figure 9. 9. hence. The printer tone curve is generally sensed at 7 to 10 different area coverages.Internal Process Controls 505 measurement-process-actuation cycle. Also. . In addition.21.19 (right). let us consider these cases separately as illustrated in Figures 9. noise in the measurements must be ﬁltered out to obtain the ﬁnal best estimate of the actual TRC. is the most widely used control system for linearizing tone levels in many digital printers. Let the Linearizing curve of DMA on photoconductor or E from paper from which TRC curve for single separation is determined 100% DMA on Photoconductor or E from paper normalized to a maximum value of 100% or 255 0 E B D A E 100% * C The TRC of DMA on photoconductor or ΔE from paper Area coverage scaled to a maximum value of 100% or 255 FIGURE 9. In these ﬁgures. the use of level 3 control can easily overcome such barriers and. If the reconstructed TRC has values available at all area coverages and is monotonic. thanks to levels 1 and 2 controls. Correcting one gray level at a time using process actuators has the risk of affecting other gray levels if the actuators are not simultaneously optimized for all levels. To demonstrate how the ‘‘static inverse’’ works.21. Whereas. the best estimate of the reconstructed TRC that passes through all the measurements is shown schematically by the continuous solid curve.9. every possible area coverage that can be printed must be measured.20 and 9. which is consistent with the fact that these points always remain ﬁxed. measurements can be carried out at different area coverages leading to a higher bandwidth tone correction with control of the entire tone curve after sampling all area coverage levels.20 and 9.1 STATIC TRC INVERSION PROCESS To completely measure the TRC. A curve ﬁtting algorithm [27] can be used to reconstruct the TRC for all area coverages between 0 and 255. the curve is forced to pass through 0% and 100% area coverage.20 Schematic diagram illustrating graphically the static TRC inversion process shown for one measurement point at area coverage A when the measured tone value is smaller than the reference tone value. complexity and the cost of implementation may prevent the use of such control loops.

to ﬁnd the tone value on the inverse TRC for area coverage A. For example. While the process of constructing ‘‘static inverse TRC’’ looks simple. This process is repeated until all the input area coverages have tone values on the inverse TRC. If not.21) until it intercepts the actual TRC. Since the real goal of level 3 controller is to linearize the tone response on paper to control the appearance of images on paper. paper-based measurements may be more desirable than those on the photoconductor. The output tone value at this intercept is indicated by C (this is also equal to B). In a digital printing system. E becomes the tone value on the inverse TRC at area coverage A. This type of inverse TRC has beneﬁts for full range of tone control from 0% to 100%. since every pixel in the image that contains input contone values is processed through the inverse TRC of the level 3 controller. draw a horizontal line (toward right in Figure 9. The rounding of values between 0 and 255 can also introduce errors which can further sacriﬁce inversion accuracy. the ﬁnal curve should be smooth and free of undesirable curvatures or artifacts introduced during the inversion process. The area coverage selection for paper-based measurements may be different from those used for measurements on the photoconductor. The tone value on the reference TRC is calculated by simply drawing a vertical line until it intercepts the reference curve. Now. whose input area coverage (or tone value) will be equal to E. which can induce undesirably high-frequency structure in the inverted TRC. image quality defects or image noise in the output images may occur. many practical considerations require attention. the measurement noise could be different for low area coverage (highlight) patches as compared to high area coverage (shadow) patches.506 Control of Color Imaging Systems: Analysis and Design Linearizing curve of DMA on photoconductor or ΔE from paper from which TRC curve for single separation is determined 100% D DMA on Photoconductor or ΔE from paper normalized to a maximum value of 100% or 255 0 C B E The TRC of DMA on photoconductor or ΔE from paper 100% E A Area coverage scaled to a maximum value of 100% or 255 FIGURE 9. Thus.21 Schematic diagram illustrating graphically the static TRC inversion process shown for one measurement point at area coverage A when the measured tone value is greater than the reference tone value. input area coverage A at which the measured tone value on the TRC is equal to D and the output tone value on the reference TRC is calculated as B. We will show later how an optimal . and these are not acceptable for high-quality color printing.20 and toward left in Figure 9.

00 77.1 with end points ﬁxed at gray levels 0 and 255 for all three reference TRCs.55 units with levels 1 and 2 controls enabled. which can be used to accurately construct the inverse tone response curve in the absence of the monotonicity condition.02 190.00 45. Figure 9.98 128.94 244. Otherwise. Construct the complete inverse TRC using static inversion process described in Section 9.76 64.24 211.Internal Process Controls 507 approach can be used to ﬁnd the right area coverages for measurements.9. and (xd)3.49 172.00 102.22.1 below.00 255. In Section 9.28 167.73 75.9. They are shown in columns (xd)1.76 128.00 51. Three patch DMA controls were used in the level 2 loop.1 TRC and Reference Samples at Discrete Input Gray Levels Number of Patches 0 1 2 3 4 5 6 7 8 9 10 U0 0.00 x 0.77 88.8 Let the vector U0 contain the input patch gray levels (or area coverages) for controlling DMA on the photoconductor for the magenta separation shown in Table 9.00 26.00 204.00 (xd)3 0.00 255. b.00 47.00 230. the static inverse methods described above will not work.87 251. This data was produced for a maximum DMA value of 0.89 255.2 we show a control-based inversion technique.00 179.00 204.17 62. In all these cases.0384 0.23 shows three plots with inverted TRCs for each reference TRCs.00 .11 255.00 179.34 218. Example 9.3738] DMA units. Three reference TRCs are sampled at the gray levels contained in the U0 vector.37 26.00 51. The shape of the reference TRCs for all gray levels are shown in Figure 9.85 82.00 230.1301 0. The vector x contains the normalized values of the measurements corresponding to each gray level. The smooth function is then inverted around the reference TRC using the static inversion technique described above.00 153.00 8. TABLE 9.35 192.31 203.00 102. a.80 91.92 242. SOLUTION Case (a): The data shown in vector x are ﬁrst interpolated are (see Chapter 6) to form a smooth curve for each gray levels.70 133.00 128. Show how the gray-level image appears when the image pixels are processed with and without the inverted TRCs obtained with each reference. The targets were set to [0.99 234.61 161.00 128. we need to ensure monotonicity in the measured (or reconstructed) tone response. (xd)2.00 153.20 43.00 (xd)2 0.14 255.00 (xd)1 0.00 26.90 197.00 77.

The cyan sweep when processed through the printer model (Chapter 10) without any level 3 correction is .508 300 Control of Color Imaging Systems: Analysis and Design (xd)3 250 (xd)2 Output gray levels 200 150 100 50 0 (xd)1 0 50 100 150 200 Input gray levels 250 300 FIGURE 9.23 Inverse TRCs for 256 input gray levels shown for the three reference TRCs.22 Reference TRCs for 256 input gray levels.24a through c. Case (b): In Figure 9. 300 DMA normalized to 255 250 200 150 100 50 0 0 50 100 150 200 Gray levels 300 DMA normalized to 255 250 200 150 100 50 0 0 50 100 150 200 Gray levels 250 300 250 300 300 DMA normalized to 255 250 200 150 100 50 0 0 50 100 150 200 Gray levels 250 300 Reference TRC Samples TRC Inverse TRC Reference TRC Samples TRC Inverse TRC Reference TRC Samples TRC Inverse TRC FIGURE 9. the cyan sweep is displayed by processing the cyan separation through three different reference TRCs.

as expected by the TRC linearization process. and : where x(k) ¼ 6 : 7. and (c) reference TRC 3.26 contains images with and without level 3 compensation.2): .2 CONTROL-BASED TRC INVERSION PROCESS A control-based TRC inversion process is made possible via a state-space representation of the tone reproduction at each of the measurement patch area coverages. shown in Figure 9.25 Circular cyan sweeps through the printer model without any inversion. Images shown on the right are processed through level 3 TRCs and the printer model.24. No printer model is involved in this simulation. FIGURE 9.Internal Process Controls 509 FIGURE 9. (b) reference TRC 2.25. B ¼ J. C ¼ I. v(k) ¼ 6 6 7 6 7 4 : 5 4 5 : xN (k) DUN (k) D ¼ 0. 9.9. y(k) ¼ x(k).24 Circular cyan sweeps through (a) identity reference TRC. A ¼ I. If the gradient of the tone curve at each of these patch area coverages is known. They show matching with corresponding ﬁgures in Figure 9. then a state-space model with an integrator in the control loop can be written as follows: x(k þ 1) ¼ Ax(k) þ Bv(k) y(k) ¼ Cx(k) þ Dv(k) 2 (9:48) 3 2 3 DU1 (k) x1 (k) 6 x2 (k) 7 6 DU2 (k) 7 6 7 6 7 7. The controller uses a gain matrix and an integrator modeled by the following equation (Section 9. Figure 9.

. e(k) ¼ xd À x(k). . U(k) ¼ U0 þ u(k) u(k þ 1) ¼ u(k) þ v(k) v(k) ¼ þKe(k) (9:49) 0 0 0 where the vector. . the error vector. . U2 .23. .26 Circular cyan sweeps through the inverted TRC obtained for reference TRCs and then through the printer model with output TRC shown in Figure 9. for the reference TRC at the patch area coverages. U1 .r2.24. rN.510 Control of Color Imaging Systems: Analysis and Design (a) Identity reference TRC (b) Reference TRC #2 (c) Reference TRC #3 FIGURE 9. . and xd is a column vector containing the desired tone values r1. . contains the patch area coverages. U0. The images on the right should match the images shown in Figure 9. . That is. UN . .

. meaning the dynamic control loop will have a SISO form. e(k) ¼ xd À x(k).49) will be diagonal.Internal Process Controls 511 2 6 07 6 U2 7 7 6 7 6 U0 ¼ 6 : 7 and 7 6 6 : 7 5 4 0 UN 0 U1 3 2 6 7 6 r2 7 6 7 6 7 xd ¼ 6 : 7 6 7 6 : 7 4 5 rN r1 3 (9:50) Also. . . U0. development of each patch can be regarded independent of the other patch. The vector x contains the normalized values of the measurements. That is. in this case. This independence is exactly the reason why a nonmonotonic TRC can be inverted using control-based approach. . Once the level 3 loop converges to a stable state with a near zero error vector. contain the input patch gray levels (or area coverages) for controlling DMA on the photoconductor for cyan separation shown in Table 9. This assumption holds good.. provided the level 2 actuators do not change (i. preserving the time hierarchy described earlier in the chapter. J. 7 6 4 5 . the inverse tone values can also be obtained from the same vector.1. This smooth curve represents the ﬁnal inverse TRC used to process the pixels in the images. the Jacobian matrix is written as 2 @x1 3 0 ÁÁÁ 0 0 @U1 6 7 6 7 @x 6 0 @U20 Á Á Á 0 7 6 7 2 7 (9:51) J¼6 . . U(k). Example 9. the Jacobian matrix. the vector xd contains the tone values of the desired reference TRC. U2 . 0 0 .e. . Since the measured tone values are present in the control vector. the measured tone values converge to the reference tone values. The measurement-process-actuation interval for level 3 control is assumed quite high as compared with level 1 and 2 controls. that is. . The derivatives are obtained at the patch area coverages for nominal operating conditions. . @x . each reference tone value is controlled independent of the others. Both vectors are shown as a function of the patch gray level. . A smooth curve is constructed to pass through the tone values in the control vector U(k) and the end points (0 and 255).. . contains the ﬁrst derivatives of the tone values at each of the patch area coverages and the structure of the matrix is diagonal because during level 3 control each patch tone value can be varied without interacting with the other patches. UN . 6 . For N number of level 3 control patches with area coverages 0 0 0 U1 .9 Let the vector. . The approach is described below with a numerical example. 7 . have fully settled) while level 3 control loop is running. . 7 6 . @UN 0 N Since the Jacobian is diagonal for level 3 control. the gain matrix (Equation 9.

c. (b) Inverse TRCs for 256 input gray levels shown for reference TRC 2.e.29b. 250 200 Gray levels 150 100 50 0 (a) DMA normalized to 255 300 250 200 150 100 50 0 Reference TRC Samples TRC Inverse TRC 1 2 3 4 5 6 7 Iteration # 8 9 10 (b) 0 50 100 150 200 Gray levels 250 300 FIGURE 9. 2.27 (a) Convergence plot (poles ¼ 0. Corresponding full TRCs are shown in Figures 9.27a. b. U(k)) iterations.27b. Plot them with respect to iteration number. U(k)) iterations. .e. . d. Use Equation 9. .512 Control of Color Imaging Systems: Analysis and Design a. 9.45 to ﬁnd the gain matrix. and 9. 2. 9.4) for the inverse tone values. k.28a. . 9. corresponding to the patch gray levels. .28 (a) Convergence plot (poles ¼ 0. (b) Inverse TRCs for 256 input gray levels shown for reference TRC 1. are shown in these ﬁgures corresponding to each patch gray level. Using end points 0 and 255 and a suitable curve ﬁtting algorithm complete the full inverse TRC. . Complete steps a through c for the new reference tone values. 250 200 Gray levels 150 100 50 0 (a) DMA normalized to 255 300 250 200 150 100 50 0 0 (b) 50 100 150 200 Gray levels 250 300 Reference TRC Samples TRC Inverse TRC 1 2 3 4 5 6 7 Iteration # 8 9 10 FIGURE 9. k ¼ 1.. and the normalized values of the measurements. respectively. and 9. Find the derivatives at the patch values and the gain matrix for pole values equal to 0. Mark the values of the reference vector.4. SOLUTION Inverted tone values for (i. Plot the inverted TRC.4) for the inverse tone values. Reference vector. . . k ¼ 1. k. 10 are plotted in Figures 9. xd. and the normalized values of the measurements.29a for three different reference TRCs as a function of the iteration number. x.. .28b. xd. Obtain the inverted tone values for (i. x.

the dead beat response can be achieved by making the closed-loop transfer function have poles at the origin (i.9.49. the theory is developed only for a discrete system. In other words. That is. ^ For the SISO case (e.2).g. . the target is reached in the fewest number of time steps. Consider the state variable description of Equation 9. the closed^ loop system matrix B will have all eigenvalues equal to zero. In the linear discrete case such as for level 1.e. 9. Thus.10 DEAD BEAT RESPONSE Dead beat response is the response of a control loop in which the error is dead in one beat.. In this section we will determine the minimum number of steps required to achieve dead beat response for the both SISO and MIMO systems.48 and the feedback equation with gain matrix of Equation 9. Section 9.29 (a) Convergence plot (poles ¼ 0. There is no such thing as a dead beat response for continuous control systems. Therefore. This implies that the desired response of the control system is obtained in a minimum number of measurement-process-actuation update steps [13].4) for the inverse tone values. The closed-loop state Equation is equal to x(k þ 1) ¼ (I À BK)x(k) þ BKxd ^ ¼ Bx(k) þ BKxd From Chapter 4.. Let l1 ¼ l2 ¼ Á Á Á ¼ 0. for a SISO system.52) will be at zero for deadbeat control. we can write the solution to the above equation as ^ x(k) ¼ Bk x0 þ BKxd (9:53) (9:52) We know that the feedback control law that assigns all the closed-loop poles to the origin is a deadbeat control law. 2. (b) Inverse TRCs for 256 input gray levels shown for reference TRC 3. Hence. zero). the B matrix will be diagonal with all À1 elements zero for deadbeat control and K ¼ B . or 3 controls.Internal Process Controls 300 DMA normalized to 255 250 Gray levels 200 150 100 50 0 (a) 1 2 3 4 5 6 7 Iteration # 8 9 10 (b) 300 250 200 150 100 50 0 0 50 100 150 200 Gray levels 513 Reference TRC Samples TRC Inverse TRC 250 300 FIGURE 9. all eigenvalues of the closed-loop system (Equation 9.

theoretically speaking.54 that the deadbeat control law can provide control in one measurement-process-actuation cycle. noise and uncertainties will make it difﬁcult to achieve deadbeat. if B is diagonalizable to pseudo-diagonal form such as the Jordan canonical form. Jordan) form (Chapter 3). This implies that for a four-dimensional ^ (4-D) system. it is possible to reach deadbeat in one measurement-process-actuation cycle. ^ if l1 ¼ l2 ¼ l3 ¼ l4 are eigenvalues of matrix B. . x(4) ¼ xd. the matrix B is diagonalizable.. ^ the B matrix can be written as follows: ^ M À1 BM ¼ J Substituting Equation 9.53 x(k) ¼ MJ k M À1 x0 þ BKxd (9:56) (9:55) ^ To make this problem simple. then from linear algebra [28].57 0 60 1 6 J ¼4 0 0 2 1 0 0 0 0 1 0 0 3 2 0 0 07 2 60 7. Since there will be repeated eigenvalues during deadbeat control. all elements of the closed-loop ^ system matrix B will not be equal to zero and is not diagonal. J ¼ 6 40 05 0 0 0 0 0 0 0 0 0 0 3 2 1 0 6 07 7. let us consider that the B matrix is of size 4 Â 4. Jk can be expressed as follows: 2 k 6 6 6 J ¼6 0 6 40 0 lk 1 kÀ1 kl1 (k À 1) lk 1 0 lkÀ2 1 2! (k À 2) (k lk 1 0 0 kÀ1 kl1 lkÀ3 1 3! lkÀ2 1 À 1) 2! kÀ1 kl1 lk 1 3 7 7 7 7 7 5 (9:57) when k ¼ 1. However. from Equation 9. as in level 1 and 2 control. J ¼ 6 40 15 0 0 0 0 0 0 1 0 0 0 3 2 0 0 17 3 60 7. Hence.e. we see that the states. Now. a minimum of four measurement-process-actuation cycles are ^ required to reach the ﬁnal desired values.514 Control of Color Imaging Systems: Analysis and Design x(1) ¼ xd (9:54) It can be seen from Equation 9. For a MIMO system. Loops must be designed away from the deadbeat response. To determine the number of measurement-process-actuation cycles required to achieve deadbeat con^ trol. in printers. and J 4 ¼ 6 0 40 05 0 0 0 0 0 0 0 0 0 0 3 0 07 7 (9:58) 05 0 For k ¼ 4. the matrix B has to be expressed in diagonal or pseudo-diagonal (i. in practice.55 in Equation 9. In most cases.

Internal Process Controls 515 9. tribo. the more toner is attracted to the photoconductor (Equations 10. developer age. the rate limitation in dispense and consumption of toner as images are rendered at production speed. Changing the TC in the developer housing can affect the lightness or darkness of a rendered image. These factors are referred to as sensor noise.13. and.11. the mixture is picked up by a developer roll. the greater attractive force between the electrostatic latent image and the toner particles causes the toner particles to transfer from the carrier granules and adhere to the electrostatic latent image. and some of them can be carefully measured. In some of the multicolor electrophotographic print engines. 9. that is.31). In this section.2.g. Developer granule size.1 OPEN-LOOP TC MODEL TC is the ratio of toner mass to the carrier mass. the toner mass (tm) at cycle k can be modeled by the following difference equation tm (k þ 1) ¼ tm (k) þ md (k À m) À mi (k) (9:59) . The toner particles and carrier granules are selected such that the toner particles acquire the appropriate charge during mixing. This mixture is transported to the development zone where the toner particles are presented to the latent image on the photoconductor. For instance.6. In simple terms. the temperature and RH affect the TC measurements. possibly. the higher the concentration of toner in the housing. toners and carriers are mixed in the sump at a speciﬁed ratio known as the TC (Section 10. The concentration of toner in the development housing has an effect on the amount of toner attracted to charged portions of the photoconductor. metered using a trim bar to achieve a uniform thickness with toner particles alone and transported to the development zone. estimated (e. When they are brought into contact with the charged photoconductive surface. The design of such a feedback system could be inherently simple if we didn’t have the complexity of the transport delay in dispensing the toner to the developer housing. As the TC grows higher. Hence. the magnetic reluctance associated with magnetic carrier particles in the developer housing. The magnetic reluctance signal is used to measure the TC in the developer. In Section 9.32]). toner age [31..30 and 10. we minimize the number of factors affecting the developability. TC can be controlled by controlling the rate at which toner from the toner supply is delivered to the developer housing. such as ferromagnetic granules. and isolated from the actual signal depending on the complexity and cost associated with the system. by maintaining TC constant at a set point.11 TC CONTROL LOOP In a typical two-component development housing. and the toner dispense rate is used as the actuator [35]. for example. the average spacing between the carrier beads gets larger and the reluctance becomes lower. we show a systematic way to realize a feedback system with a transport delay using state feedback control approaches. The toner particles consist of dyed or colored thermoplastic powder that is mixed with coarser carrier granules.3). the TC can be used as one of the xerographic process actuators [16] in addition to those used in level 2 controls as described in Section 9. TC sensors [29–36] are used to sense.

md (k). then we have tc (k) ¼ Rewriting Equation 9. This assumption is unrealistic for most of the practical development systems. The mi(k) is calculated by using the area coverage a(k) of the image. The mass dispensed is calculated by using the duty cycle which will be derived by processing the measured TC signal inside the feedback controller.59 in terms of TC tc (k þ 1) ¼ tc (k) þ g[u(k À m) À v(k)] (9:63) tm (k) mc (9:62) (9:61) where u(k) is the dispensed mass. After the transformation we get .516 Control of Color Imaging Systems: Analysis and Design where tm(k) stands for the mass of the toner at cycle k in grams md(k) is the mass of toner dispensed from the dispenser at cycle k in grams mi(k) is the mass of toner which is used for the image (based on the consumption proﬁle at cycle k in grams) m is the transport delay in toner dispense (in number of cycles) that includes the delay in the mixing system The index k denotes the iteration number (or measurement-process-actuation cycle). a pixel counter is included in many modern printing systems that can be used to calculate the total area coverage of the image in which the toner will be developed. It is given by mi (k) ¼ a(k) Â Rmax Â t From the deﬁnition of TC. if mc is the carrier mass. mi (k). we assume that the TC cycle is synchronous with the measurementprocess-actuation cycle of the level 2 control system. and g is a scale factor given by g¼ 1 mc (9:64) The Equation 9. and v(k) is the developed mass. md (k) ¼ d(k) Â Rmax Â t (9:60) where Rmax is the maximum dispense rate at which the toner can be dispensed from the toner reservoir t is the measurement-process-actuation period in seconds d(k) is the duty cycle at cycle k Generally.63 can be transformed to the z-domain by applying the z-transform. For simplicity.

. the open-loop TC system becomes unstable if the developed mass v(k) is zero. Note that the TC sensor model is ignored in this design.30. Figure 9. The parameter ~(z) ¼ u(z)=t represents the rate of toner u dispense and t is the TC measurement-process-actuation interval in seconds.Internal Process Controls 517 v(z) Gp(z) _ u(z) G(z) = Gp(z)z–μ + tc(z) FIGURE 9. and the error signal.e. 9. with Kp ¼ 0) will give stable performance provided the gains are designed correctly.30 Open-loop transfer function of TC control system. then a simple proportional integral (PI) controller can be used to stabilize the system.31 shows the PI controller in the z-domain. Clearly. if we ignore the time delay. w(z). es(z). The system of Equation 9. It is important to note that a simple integral control (i. are related by the following transfer function: Kp td(z) + + – w(z) z z–1 + PI controller Ki Gpi(z) v(z) Gp(z) + _ tc(z) es(z) + + τ ~ u (z) G(z) = Gp(z)z–μ u(z) Plant FIGURE 9.11. For now.31 TC feedback system with PI controller. The intermediate quantity. The block diagram of the open-loop TC system of Equation 9.66 is shown in Figure 9.66 contains an integrator and a time delay. this condition occurs when the developer housing is being cycled to ﬁll the toner without printing images. . tc (z) ¼ or g ðzÀm u(z) À v(z)Þ zÀ1 (9:65) tc (z) ¼ Gp (z)ðzÀm u(z) À v(z)Þ (9:66) where Gp(z) ¼ g=(z À 1).2 DESIGN OF A TC CONTROL LOOP USING A PI CONTROLLER Pole-placement design is applied to bring the TC from the initial state to the desired state in few cycles.

. For simplicity. A¼ .e.63 can be written as tc (k þ 1) ¼ tc (k) þ gu(k) Now. we get the following difference equation: w(k þ 1) ¼ w(k) þ td (k þ 1) À tc (k þ 1) (9:68) For a TC system with no dispenser lag (i. E¼ .72 can then be written in matrix form as Â Ã Àes (k) u(k) ¼ Àt Kp ÀKi w(k) ! (9:73) (9:72) . let us assume a constant TC target. Equations 9..67 and taking the inverse z-transform. the toner will start to lose its charging property.70 can be written in more general matrix difference (state space) form for a single-input system as x(k þ 1) ¼ Ax(k) þ Bu(k) þ Er(k þ 1) (9:71) ! ! ! ! 1 0 g 0 tc (k) .e. the amount of dispensed toner can be expressed as a function of the controller gains: À Á u(k) ¼ t Kp es (k) þ Ki w(k) Equation 9. For a constant TC target. B¼ . td(k þ 1) ¼ td(k) ¼ constant). Equation 9.69 are arranged in matrix form as follows: ! 1 tc (k þ 1) ¼ w(k þ 1) À1 0 1 ! ! ! ! tc (k) g 0 t (k þ 1) þ u(k) þ w(k) Àg 1 d (9:70) (9:69) Equation 9.66 in statespace form as follows: Substituting the error signal es(z) ¼ td(z) À tc(z) in Equation 9. and r(k) ¼ td(k). m ¼ 0) and no consumption (i. This is called toner aging [31]. v(k) ¼ 0). there is a signiﬁcant difference in toner transfer to the paper. As a result of this.e.518 Control of Color Imaging Systems: Analysis and Design w(z) ¼ z es (z) zÀ1 (9:67) The gains of the PI controller are designed by ﬁrst expressing Equation 9. Due to the repeated mechanical stresses during the mixing process. In some systems.. with x(k) ¼ w(k) À1 1 Àg 1 The desired TC value is considered constant (i. Fresh toner has higher additive coverage whereas the aged toner has almost no additives on the toner surface.68 and 9. the TC target is adjusted at a lower rate as a function of toner age in order to achieve better control for low area coverage images. which has a negative impact on development and transfer performance.

32 shows these equations in block diagram form after the equations are transformed to z-domain. Ac ¼ A À BKC in Equation 9. F¼ . A systematic approach is required to design the controller gains.71 and assuming r(k þ 1) ¼ r(k).2 or 5.32. The controller gains can be designed using pole placement or linear quadratic regulator (LQR) techniques (Section 5. Recall that F r(z) E + + z–1I x(z) A + y(z) A B u(z) –K FIGURE 9.75 represent the TC control system in state variable form and Equation 9. Figure 9.31 The vector y(k) can be written in following form: y(k) ¼ ! Àes (k) 1 ¼ w(k) 0 0 1 ! ! ! À1 tc (k) þ td (k) w(k) 0 (9:75) (9:74) ¼ Cx(k) þ Fr(k) where C ¼ ! ! ! 1 0 À1 t (k) . following closed-loop state equation is obtained: x(k þ 1) ¼ (A À BKC)x(k) þ (E À BKF)r(k) (9:76) The matrix.Internal Process Controls 519 or in a more compact form as u(k) ¼ ÀKy(k) where K ¼ [tKp ÀtKi] is the gain matrix y(k) is the output vector for the system of Figure 9.76 represents the closed-loop system matrix of Figure 9. Hence. Substituting Equation 9.71 and 9. and r(k) ¼ td (k) w(k) 0 1 0 Equations 9.3) [13]. the control system is SIMO (single input multiple output). the closed-loop characteristic matrix is zI À Ac ¼ 0. Below we show the poleplacement design.32 Block diagram of the feedback system with PI controller.74 in Equation 9. x(k) ¼ c .74 the output feedback equation. Note that although the original system was SISO. .

and K ¼ tKp ÀtKi 0 1 0 tc (k) (9:78) (9:77b) (9:77a) SOLUTION We use pole-placement design for SISO system discussed in Section 5.65. Design the gain matrix for placing the poles at (a) l1 ¼ 0. Àg w(k) À1 1 1 ! ! Â Ã 1 0 À1 C¼ . A¼ . E¼ . Case (a): Since TC target is to be calculated as a percentage and the carrier mass ¼ 800 g. In other words.2. F¼ .77. we get x(k þ 1) ¼ (A À BK)x(k) À BKFr þ Er (9:81) (9:80) .6. B¼ . r(k) ¼ td (k). l2 ¼ 0. Assume TC target in percent.520 Control of Color Imaging Systems: Analysis and Design the gain matrix K places the poles (or roots) of the characteristic equation. See Example 9. l2 ¼ 0. we can write g ¼ 100=800 ¼ 1=8. Example 9. Carrier mass ¼ 800 g.10 State equations for a closed-loop TC control system are given below.2. K is chosen by assigning the closed-loop poles.25 and (b) l1 ¼ 0. The state equation is x(k þ 1) ¼ Ax(k) þ Bu(k) þ Er(k þ 1) And the output equation is y(k) ¼ Cx(k) þ Fr(k) The actuator with output linear feedback is given by u(k) ¼ ÀKy(k) where x(k) ¼ ! ! ! ! 1 0 g 0 . The open-loop TC equation is given by " # ! ! 1 1 0 0 8 x(k) þ r x(k þ 1) ¼ u(k) þ À1 1 1 À1 8 (9:79) where r is the constant value of TC expressed in percent. The actuator equation can be written as follows: u(k) ¼ ÀKy(k) ¼ ÀK[Cx(k) þ Fr(k)] ¼ ÀKCx(k) À KFr ¼ ÀKx(k) À KFr Substituting Equation 9.80 in Equation 9.2.10 for more details.

ﬁnd the eigenvalues of (A À BK) using above gain vector. l2 ¼ 0.30) is given by T ¼ QW ¼ ½ B b AB 1 1 ! 1 1 1 ¼ 0 8 À1 1 À2 ! ! 1 À1 À2 1 ¼ 1 0 8 0 1 À1 ! (9:85) The gain vector K (Equation 5.19.4.121) Q ¼ ½ B AB ¼ 1 1 1 8 À1 À2 ! (9:82) is of full rank.84. Hence both states can be controlled using a single actuator.83. l2 ¼ 0. designing K to place the eigenvalues of the closed-loop matrix (A À BK) at locations l1 ¼ 0. The characteristic polynomial of the open-loop system is l P(l) ¼ jlI À Aj ¼ 1 0 ¼ (l À 1)(l À 1) ¼ l2 À 2l þ 1 ¼ 0 lÀ 1 (9:83) Comparing the open-loop characteristic polynomial. u(k).39. Hence the gain matrix K is given as . The characteristic polynomial of the desired closed-loop system is Pc (l) ¼ (l À 0:2)(l À 0:25) ¼ l2 À 0:45l þ 0:05 (9:84) Comparing the closed-loop characteristic polynomial.65.45.2. b2 ¼ 1.25 will lead to a stable solution provided the controllability matrix (Equation 4. a1 ¼ À0.25.6.Internal Process Controls 521 Now. b1 ¼ À2.05. Case (b): Applying similar pole-placement techniques for SISO design. Equation 9. The transformation T is the same as in Equation 9. The transformation T (Equation 5. with Equation 5. gain matrix is calculated to place the closed loop eigenvalues at l1 ¼ 0. and a2 ¼ 0. with Equation 5. Equation 9.21.85. l2 ¼ 0.2. which we can see by inspection (the two columns of Q are independent) is the case which is full. The characteristic polynomial of the desired closed-loop system is Pc (l) ¼ (l À 0:6)(l À 0:65) ¼ l2 À 1:25l þ 0:39 (9:87) Hence a1 ¼ À1. and a2 ¼ 0.32) becomes " K ¼ ½a2 À b2 a1 À b1 T À1 ¼ ½À0:95 1:55 À1 8 0 1 8 #À1 ¼ [7:6 À4:8] (9:86) À1 8 Check: To know whether our computation is correct. It should give l1 ¼ 0. To ﬁnd the gains let us use the procedure outlined in Example 5.25.

Reference [39] shows a design method of the Smith predictor with varying time delay. Figure 9. Here we present the design of a Smith predictor for a ﬁxed time delay. Hence.65] 5 10 15 20 25 30 35 40 45 50 Iteration # FIGURE 9.2 0. It can induce oscillations. There are many methods and techniques for the analysis and control of dynamic systems in the presence of time delay [37]. v(k) ¼ 0.11.2 0. Actuator saturation can make the TC system nonlinear. K ¼ ½a2 À b2 a1 À b1 T À1 9. The desired eigenvalues can be adjusted to reduce the magnitude if it violates the actuation limits. For systems with a varying time delay. Assuming zero consumption.6 0.4 3.6 0.25] Poles at [0. when the actuator.31.4 Toner concentration in % Control of Color Imaging Systems: Analysis and Design 8 7 6 5 4 3 2 1 0 –1 0 4. dispenser rate rails at its maximum limit). l2 ¼ 0.3 DESIGN OF A TC CONTROL LOOP WITH USING A PI CONTROLLER A TIME DELAY For a TC system with a dispenser lag.25] Poles at [0.6.2 3 0 5 10 15 20 25 30 35 40 45 50 Iteration # Dispensed mass in grams Poles at [0. corresponding actuator signal (right). and G(z) ¼ Gp(z)zÀm is the transfer function . underestimating or overestimating the time delay signiﬁcantly degrades the control quality.. This delay can affect the stability and control of the TC system.33 shows the TC convergence plot and the corresponding actuator curve for a step change in TC from 3% to 4% for case (a) and (b).e.2 4 3. It can also lead to chaotic behavior when the maximum dispenser rate is limited by the admix of the material package (i. that is. Actuator limits have not been considered in this calculation.65] Poles at [0.33 TC convergence plot for a step input (left).522 4. À1 1 8 8 ¼ ½À0:61 0:75 0 À1 8 " #À1 ¼ [4:88 À1:12] (9:88) For this gain vector.6 3. if Gpi(z) is the transfer function of the PI controller shown in Figure 9. clearly the eigenvalues of (A À BK) are l1 ¼ 0. The Smith predictor-based control scheme is the most commonly used technique in industrial systems [38–40].8 3. and eventually destabilize the system even though the closed-loop poles are assigned to ensure stability. the time delay component m 6¼ 0. All of these aspects motivate the need to compensate for the delay effects.65. the actuator signal shows a very high dispense magnitude to meet the demand.

To implement the Smith predictor. we can write es (z) ¼ e(z) À ð1 À zÀm ÞGp (z)u(z) z ~(z) ¼ Kp es (z) þ Ki es (z) u zÀ1 u(z) ¼ t~(z) u (9:90) (9:91) (9:92) A useful form to implement in the computer is the difference equation form which contains a sequence of numbers iterated from each measurement-process-update cycle. is replaced by g=(z À 1) in Equation 9. After this. m cycles.34 is converted to difference equations as follows. is governed by the transfer function of the controller-plus-undelayed plant while the actual response of the controller is subjected to a time delay of m cycles. Gp(z).96 are recursive. let us write the important equations of the system with Smith predictor and PI controller in the z-domain.95 with known desired TC and measured TC values for a u known dispense process. we compute es(k) and ~(k) from Equations 9. For computation. From Figures 9. as in Section 9. the transfer function. then the closed-loop transfer function is given by Gc (z) ¼ Gpi (z)Gp (z)zÀm 1 þ Gpi (z)Gp (z) (9:89) Equation 9.89 implies that the closed-loop system response of a PI controller designed by ignoring time delay. input to the PI controller is modiﬁed by a feedback term (1 À zÀm) Gp(z) as shown in Figure 9.34.93 . First. the block diagram of Figure 9.31 and 9. we ﬁrst start with Equations 9.96 and 9. of the open-loop TC system with a dispenser lag.11.Internal Process Controls 523 1–z–μ td(z) + – e(z) + es(z) – Gpi(z) Gp(z) tc(z) u(z) Gp(z)z–μ FIGURE 9. To do this.93 through 9. which is easy to interpret.2. With the Smith predictor.34.90 and inverse z-transform is applied on the resulting equation to yield the difference equation form: es (k) ¼ es (k À 1) þ e(k) À e(k À 1) À gu(k À 1) þ gu(k À m À 1) ~(k) ¼ ~(k À 1) þ (Kp þ Ki )es (k) À Kp es (k À 1) u u u(k) ¼ t~(k) u e(k) ¼ td (k) À tc (k) (9:93) (9:94) (9:95) (9:96) Equations 9.34 Block diagram of the feedback system with Smith predictor.

94.35 TC and actuator response for a step change to TC target with time delay of three cycles and without Smith predictor (dashed curve: TC target input.11 Use Equation 9.4 3. with appropriate initial conditions.2 4 3.8 3.38.65 and Smith predictor Equations 9.4 Toner concentration in % 4. dispensed mass [right]).37. solid curve: TC output [left]. The Smith Without Smith predictor 1000 Dispensed mass in grams 5 10 15 20 25 30 35 40 45 50 Iteration # 500 0 –500 –1000 –1500 –2000 –2500 0 5 10 15 20 25 30 35 40 45 50 Iteration # 300 Toner concentration in % 250 200 150 100 50 0 –50 –100 0 FIGURE 9. SOLUTION We show the results in Figures 9.35 and 9. With Smith predictor 7 Dispensed mass in grams 5 10 15 20 25 30 35 40 45 50 Iteration # 6 5 4 3 2 1 0 0 5 10 15 20 25 30 35 40 45 Iteration # 50 4.2 3 2. The dispense rate ~(k) is u used to actuate the dispenser.36 TC and actuator response for a step change to TC target with time delay of 3 cycles and with Smith predictor (dashed curve: TC target input. respectively. Clearly the feedback system is unstable for delay of 3 and 10 cycles as shown in Figures 9.8 0 –1 FIGURE 9.35 through 9. Assume no images are printed during the entire simulation.96 to simulate the performance of TC system with a dispenser lag of (a) 3 cycles and (b) 10 cycles. solid curve: TC output [left]. Example 9. .6 3.6.93 through 9.63 as the TC system and the PI controller with closed-loop poles l1 ¼ 0. dispensed mass [right]). l2 ¼ 0.524 Control of Color Imaging Systems: Analysis and Design and 9.

u .Internal Process Controls Without Smith predictor 2000 Dispensed mass in grams 1500 1000 500 0 525 50 Toner concentration in % 0 –50 –100 –150 –200 –250 0 5 10 15 20 25 30 35 40 45 50 Iteration # –500 0 5 10 15 20 25 30 35 40 45 50 Iteration # FIGURE 9. Proper feedforward compensation of the control input. dispensed mass [right]). the system can be overloaded with disturbance before it has time to recover. dispensed mass [right]).8 0 –1 FIGURE 9. from the pixel counter can potentially minimize the impact. solid curve: TC output [left].11.4 3.2 4 3.37 TC and actuator response for a step change to TC target with time delay of 10 cycles and without Smith predictor (dashed curve: TC target input. pixel count is used to estimate the amount of toner required for developing the images.38). if the jobs contain a mixture of low and high area coverages and the change is rapid.8 3. 9. depending on how the area coverages are distributed while printing. solid curve: TC output [left].4 Toner concentration in % 4. predictor improves the stability of the loop (Figures 9. Since images can have varying degrees of area coverage. while at the same time injects some overshoot when compared to the response without any delay as in Figure 9.38 TC and actuator response for a step change to TC target with time delay of 10 cycles and with Smith predictor (dashed curve: TC target input.4 FEEDFORWARD COMPENSATION FOR IMAGE DISTURBANCE Generally.36 and 9.2 3 2. This can induce undesirable ﬂuctuations in TC and even a loss of stability.33. Particularly.6 3. it can represent a large disturbance to the TC control system. ~. With Smith predictor 7 Dispensed mass in grams 5 10 15 20 25 30 35 40 45 50 Iteration # 6 5 4 3 2 1 0 0 5 10 15 20 25 30 35 40 45 50 Iteration # 4.

since the controller does not use an integrator. sensor). an antiwindup compensator will not be required if the TC loop were to operate under limited actuator (Section 9. and the estimated toner mass is known accurately (which is included in the feedforward gain. which is calculated for each image using the pixel counter. Optimization of the toner replenishment system and measurement system is related to the system hardware.12). can be found using following equation: v(k) ¼ image size Â toner mass Â area coverage (k) 1000 (9:97) where toner mass is the amount of toner developed in milligrams=square centimeters.526 Control of Color Imaging Systems: Analysis and Design We ﬁrst compute the area coverage from the pixel counter.91) is modiﬁed with feedforward compensation as shown below: ~(z) ¼ Kp es (z) þ u z Ki es (z) þ Kff zm pc (z) zÀ1 (9:98) 9. States are estimated from the sensor output using a state estimator. . The estimated amount of developed mass. States are deﬁned by creating an open-loop state-based model of the system.g. In this algorithm.e. As the system and each of the states is controlled. perfect cancellation for image disturbance can be achieved. Area coverage ¼ Pc. This is split into m states as in Figure 9. v(k). the time delay in the toner dispense is replaced by states.. In this section. a new state-based feedback control algorithm (i.39. Kff). plant itself).500 Â 1100 page with 10 pages being printed per one TC cycle).. Optimization of the feedback controller is a software function which is easy to be incorporated even after the hardware system is fully commissioned.94 Â 10 square centimeters for a 8. In the actual implementation of the feedforward compensation function. a set of instructions implemented via a computer code) is discussed. and (3) the closed-loop control algorithm.5 DESIGN OF TC CONTROL LOOP WITH STATE FEEDBACK CONTROLLER AND STATE ESTIMATOR To improve the overall TC performance.e. If the area coverage of each image is known m cycles ahead of actual print. (2) the TC measurement system (i. In addition to this beneﬁt.e. in grams...59 Â 27. 21. Kff. is the ratio of the total area where there is toner to the total image area. an improved dynamic performance is expected as compared with the Smith predictor implementation. and the resulting signal is advanced by m cycles to account for the dispenser lag. the following three major system components need to be optimized: (1) the open-loop toner replenishment (dispense) and mixing system (i. The system described by Equation 9. the pixel counter data is multiplied by a constant.63 has a delay of m cycles.11. and each state is estimated using measurements from the TC sensor. Image size is the size of the image (e. Formations of these states can be understood when the system equation is expressed in terms of z-transforms. The control input (Equation 9.

These vectors and matrices are (9:100) . .39. . which in turn is related to the toner mass. the state equations can be written as x1 (k þ 1) ¼ x1 (k) þ gu(k) x2 (k þ 1) ¼ x1 (k) . number of states will be reduced to two (one due to the mass dispensed and another due to the mass developed) as seen from the state equation below. If the time delay is zero. A 2 R(mþ1)Â(mþ1) .99. B 2 Rmþ1 . xmþ1 (k þ 1) ¼ xm (k) xmþ2 (k þ 1) ¼ xmþ2 (k) þ gv(k) and the output equation is given by y(k) ¼ tc (k) ¼ xmþ1 (k) À xmþ2 (k) (9:99b) (9:99a) It is important to note that. From the block diagram shown in Figure 9. in Equation 9. the states represent TC at a discrete delay cycle. xmþ2 (k) 2 R1 . states can be deﬁned as quantities related to the toner mass at discrete delay cycle. when the toner dispense is instantaneous. u(k) 2 R1 . Total number of states depends on the delay cycles.Internal Process Controls v(k) 527 g z–1 x1(k) u(k) g z–1 1 z x2(k) 1 z x3(k) 1 z xμ+1(k) + _ xμ+2(k) y(k) = tc(k) FIGURE 9. for the TC system. that is. In physical terms. and C 2 Rmþ1 . The above equations can be grouped into matrix form as x(k þ 1) ¼ Ax(k) þ Bu(k) xmþ2 (k þ 1) ¼ xmþ2 (k) þ gv(k) y(k) ¼ Cx(k) À xmþ2 (k) where x(k) 2 Rmþ1 .39 Block diagram of open-loop TC model with time delay.

.528 Control of Color Imaging Systems: Analysis and Design mi(k) Feedforward compensator xtarget Target state decomposer + _ ˆ xtarget(k) TC target td(k) Actuator saturation Open loop TC system TC output y(k) State estimator Gain matrix Kc ˆ x1 (k) ˆ x2 (k) ˆ xμ+1(k) FIGURE 9. .5 0 and 0 ÁÁÁ (9:101) This state-space equation is used to develop the state feedback estimator and state feedback controller as follows. . . with K0 as the observer gain is considered as the state estimator. x(k) ¼ 6 . 1 0 2 3 g 607 6 7 B¼6.102. 4. The states of the open-loop TC system can be estimated using a Kalman ﬁlter (an optimal observer) or a linear full-order observer (a suboptimal observer).40 controller.7 .5 .7 4. . . A full-order observer (Section 5. Block diagram of open-loop TC model with state estimator and state 2 6 x (k) 7 6 2 7 7. 0 0 1 ÁÁÁ 0 3 ÁÁÁ 07 7 7 Á Á Á 0 7. . The block diagram of the closed-loop TC control system with state estimator and state controller is shown in Figure 9. Â Ã ^(k þ 1) ¼ A^(k) þ Bu(k) þ K0 y(k) À C^(k) þ xmþ2 (k) x x x (9:102) . 4 5 .7 6. 0 0 0 .2) [41]. 7 . 6 7 .40. xmþ1 (k) C ¼ ½0 0 ÁÁÁ x1 (k) 3 2 1 0 0 1 . . 61 6 6 A ¼ 60 6 6. Equation 9.4.

102) by measuring the TC. The dynamic TC system. they are estimated using a linear observer (Equation 9.126) Â P ¼ C0 A0 C 0 :: (A0 )mþ1À1 C 0 Ã (9:105) be full rank. The matrix A À K0C is called the observer matrix. the observer and the state feedback designs can be decoupled and designed independently.84 to design the observer gain matrix.100 is a SISO system unlike our augmented system (Equation 9. Having determined the observer gain matrix and the controller gain matrix. and we designed the state feedback control law. we chose the gain matrix K0 to place the eigenvalues of the observer matrix within the unit circle so that the observer states asymptotically converge to the true state. based on the area coverage information.104. Determination of the gain matrix that accomplishes this is a pole-placement task. all the eigenvalues of the observer matrix must lie inside the unit circle in the z-domain. Equation 9.102 as follows: e(k þ 1) ¼ x(k þ 1) À ^(k þ 1) ¼ (A À K0 C)(x(k) À ^(k)) x x Since e(k) ¼ x(k) À ^(k) Equation 9. It is a simple linear transformation formed . is accessible for measurement. For the error to go to zero asymptotically. Equation 5. The number of gains is equal to the number of states that the control input u(k) affects during feedback. it is necessary that the observability matrix (Equation 4. these processing units have to be interconnected with the TC target to derive proper control input. the control input is calculated using the controller gains (row vector) and estimated states as follows: u(k) ¼ ÀK^(k) x (9:106) The system described by Equation 9. xmþ2(k þ 1) ¼ xmþ2(k) þ gv(k).75) with the PI controller.40) is designed for creating a vectorized version of the TC target so that the output from the state feedback controller can be subtracted correctly. After the states are estimated. We can use the Bass–Gura formula or Ackermann formula. x(k). The target state decomposer (shown in Figure 9. e(k) ¼ x(k) À ^(k) reaches zero steady state.100. Note that. y(k).100 and 9. Therefore. But. A difference equation x can be obtained for the error signal by using Equations 9. The gain vector K ¼ [K1 K2 Á Á Á Kmþ1] can be obtained by using the pole-placement or optimal control techniques. because of our inability to measure the system states. By the separation theorem. in order to place the poles of the observer matrix. is in the linear state-space form.103 can be written as x e(k þ 1) ¼ ð A À K0 CÞe(k) (9:104) (9:103) The observer is in the form of a feedback system with the error dynamics satisﬁed by Equation 9. with a TC sensor and estimating the feedforward signal.Internal Process Controls 529 The gain K0 controls the rate at which the error signal between the true states and the estimated states. u(k) ¼ ÀKx(k) with K ¼ [K1 K2 Á Á Á Kmþ1] under the assumption that the state vector.

Windup can decrease the overall performance of the control system and cause large overshoot. are switched off for some valid reason. the actuator will remain at its limit until the system output crosses the target value and sufﬁcient error of opposite sign is integrated to remove the windup. to provide an increasing control signal that brings the process output back to the desired target. development. etc. practical limits are set for charge voltage. There are. At the low end. over time. a phenomenon known as integrator windup occurs whenever the actuator signal is driven to its maximum limit by the controller in an effort to bring the output to its desired target. This is necessary to sustain reliable operation of production printing systems. in the TC control system. there will be a nonzero error that will continue to be integrated by the integrator and this phenomenon is known as ‘‘windup. and TC control systems use some form of integral control to reduce steady-state errors. For example. however. 3 td 6 td 7 6 7 ¼ 6 : 7. xtarget 2 Rmþ1. These short comings of the integral controllers can be minimized by including an antiwindup compensation. and large steady-state errors. several aspects of an integral controller that can potentially have ill effects on the performance of the system when the dynamic range of the actuators is bounded by some practical limits [42–48]. It can also occur in other situations like when the loop dynamics are changed abruptly. an integral controller integrates even small variations of the error signal between the sensed value and the target value. If the output does not return to the desired target. This is because.’’ As a result. exposure.12 PROCESS CONTROLS UNDER LIMITED ACTUATION In modern electrophotographic printers. the rate is set to a minimum value to avoid the small rates that cannot be provided reliably by the dispenser motor.530 Control of Color Imaging Systems: Analysis and Design by multiplying the controller gain matrix with elements arranged along the diagonal with the desired TC value. and when a sensor fails or when there is inaccurate feedforward compensation due to measurement errors. Similarly. xtarget ¼ Kc xdesired (9:107) where xdesired 2 Rmþ1. and Kc 2 R(mþ1)Â(mþ1) is the controller gain matrix with elements along the diagonal. many . Although an optimization-based antiwindup design for MIMO control system is still an active area of research and many questions remain unanswered. 6 7 4:5 td 2 xdesired Â Kc ¼ diag K1 K2 ÁÁÁ Kmþ1 Ã (9:108) 9. That is. When an integrator is used in the control loop. the maximum rate is limited by the admix of the material package. exposure raster output scanner (ROS) intensity and the development bias voltage due to cost and other considerations. dispense rate is limited to low and high values. At the high end. longer settling time.

and an antiwindup compensator. when the integrator saturates. which may happen v u when the development voltage is changing rapidly. a Smith predictor. the antiwindup feedback obtained through the high gain. The output of the integrator ~ is the input to the high gain v feedback. for example. the integrator output may become excessively large. The integration of the error signal will lead to a larger ~ which will v have no effect on the TC output since the actuator is limited to ~max. When v u ~ saturates to ~max during the course of the loop operation. During this time. To compensate for these effects. between ~min and ~max). Use actuator saturation.. u In this scheme.e. When ~ is within the actuator limits v u (i. opposite to what was happening when ~ saturates to ~max). In the absence of actuator limits.. This is achieved by inserting a high gain v u v dead zone feedback using another error signal obtained by subtracting ~max from ~. . L. If the u integrating effect continues further in the same direction.12 Consider the design of a TC feedback system with a transport delay using a state feedback estimator and controller. Ki. the actuator signal ~ will be same as ~. systems use a trial and error-based approach for compensation. as an alternate feedback structure.41 PI controller with antiwindup compensation for a TC system. Apply the area coverage disturbance shown in Figure 9. The high gain antiwindup compensator can also be applied in Figure 9. The implementation of the actuator limits would involve the use of a saturation function to generate the control signal ~. at the integrator becomes active to force the integrator windup error to zero. Example 9.41 before the integrator constant.Internal Process Controls 531 Kp z z–1 ~ v ~ u ~ umax ~ umin L ~ v ~ u τ TC system tc(z) es Ki FIGURE 9. Construct a ﬁgure of merit to compare the results.41 shows a typical antiwindup scheme used in printers for a SISO TC system. the error signal es will be nonzero.42. Design the observer gain matrix and the controller gain matrix using pole-placement techniques. Figure 9. Compare the TC response using the conventional approach with a PI controller. the integrator acts like a fast ﬁrst-order lag with a transfer function that has a single pole created by the antiwindup compensator gain L.e. a dead zone is created by disabling the high gain u feedback automatically. we need to create a situation where the sign of the error is inverted (i. which is the input to the high u gain antiwindup compensator loop. The value of the compensator gain determines how fast the integral action is pulled back from actuator saturation.

5 3 2. and noise.12 0.14 0. and area coverage disturbance. actuator saturation.1 0. The PI controller . actuator limitation.42 Area coverage disturbance shown with respect to TC update cycle.5 0 20 40 60 80 100 120 Print number 140 160 180 200 FIGURE 9.08 0.43 through 9.02 0 Area of coverage 0 20 40 60 80 100 120 140 160 180 200 FIGURE 9. 5 Smith predictor Target State feedback 4. SOLUTION The response to a step change in TC between 3% and 4% with area coverage disturbance is shown in Figures 9.16 0.04 0.06 0.43 TC Response for a step input (3%–4%) in the presence of time delay.5 4 3.532 Control of Color Imaging Systems: Analysis and Design 0.47 in the presence of time delay. area coverage disturbance.

5 3 2. actuator limitation.5 0 20 40 60 80 100 120 Print number 140 160 180 200 FIGURE 9.5 4 3. and area coverage disturbance and 1% sensor noise.Internal Process Controls 533 5 Smith predictor Larget State feedback 4.45 TC Response for a step input (3%–4%) in the presence of time delay. and area coverage disturbance and 2% sensor noise. 6 Smith predictor Target State feedback 5 4 3 2 1 0 0 20 40 60 80 100 120 Print number 140 160 180 200 FIGURE 9.44 TC Response for a step input (3%–4%) in the presence of time delay. . actuator limitation.

. and area coverage disturbance and 3% sensor noise. actuator limitation.534 Control of Color Imaging Systems: Analysis and Design 6 Smith predictor Target State feedback 5 4 3 2 1 0 0 20 40 60 80 100 120 Print number 140 160 180 200 FIGURE 9.46 TC Response for a step input (3%–4%) in the presence of time delay. and area coverage disturbance and 5% sensor noise.47 TC Response for a step input (3%–4%) in the presence of time delay. 6 Smith predictor Target State feedback 5 4 3 2 1 0 0 20 40 60 80 100 120 Print number 140 160 180 200 FIGURE 9. actuator limitation.

the overshoot due to time delay has been reduced.6 Cyan area coverage 0. we show the area coverage disturbances expected from a sample print job. This is because of the integral part of the controller.7 0. The state feedback technique works well because the output is fed back to a number of states instead of just one as in PI controller.4 0. meaning the ﬁnal TC will not be equal to what was requested (in this case 4%).6 Magenta area coverage 0.8 0.2 0. (continued ) . k 200 250 FIGURE 9. By including the Smith predictor.4 0. All states are estimated and controlled.1 0 0. In Figure 9.5 0. k 200 250 0 50 100 150 Print number. Clearly it appears as noise without any structure. but the sensor noise can give rise to increased steady-state error. the error correction is more favorable with state feedback.49 shows an 0.7 0.Internal Process Controls 535 without a Smith predictor gives an overshoot (not shown in the ﬁgures).1 0 (a) 0 50 100 150 Print number.3 0.5 0.2 0. thus.48. But the state feedback gives reduced steady-state error under those noise uncertainties.48 Area coverage proﬁle used for the TC control algorithm (obtained using a scanner from magazine samples).3 0. Figure 9.

The ﬁgure of merit.5 0.1 0 Control of Color Imaging Systems: Analysis and Design Yellow area coverage 0 50 100 150 Print number. improved steady-state error as compared with the PI controller with Smith predictor and a nonoptimized model predictive controller (not described in this book). which is deﬁned as the ratio of the area under the curve with disturbances and uncertainties to that without any such disturbance. k 200 250 (b) FIGURE 9. When the TC response is above the target value. More advanced methods such as model predictive controller can be further optimized to work similar to or better than the state feedback method for the same problem.4 0. The ﬁgure of merit in Table 9.3 0.2 shows overall improvement can be achieved with state feedback methods. . then it is subtracted from the required area.2 0.7 0.25 0. k 200 250 0. the ideal ﬁgure of merit is a step response from 3% to 4%.05 0 0 50 100 150 Print number.6 0. is calculated by measuring the area under the response curve with respect to the target.536 0. In our comparison.2 0.1 0.15 0.4 0.35 0.3 Balck area coverage 0.48 (continued) Area coverage proﬁle used for the TC control algorithm (obtained using a scanner from magazine samples).

v ¼ umax when v > umax. k 140 160 180 200 Smith predictor Target State feedback Model predictive 537 FIGURE 9.8 3. D ¼ Kp. C. The controller state-space equation with the parameter matrices. v(k) ¼ ~(k). in a SISO system. The uncompensated MIMO PI controller is shown in Figure 9. and the controller state vector x(k) is as follows: x(k þ 1) ¼ Ax(k) þ Be(k) v(k) ¼ Cx(k) þ De(k) (9:109) The saturation function can be written in mathematical form as v ¼ u. where umin and umax are the minimum and maximum allowable actuator vectors. Clearly. B.2 3 2.49 TC Response for a step input (3%–4%) in the presence of time delay. e(k) ¼ es(k). the input error vector e(k). the PI controller of Figure 9.8 0 20 40 60 80 100 120 Print number.50. B ¼ Ki. has parameters. which is obviously more complex than the MIMO level 1 system. A. and area coverage disturbance. 5% sensor noise and area coverage disturbance of Figure 9.48.4 3.6 4. v ¼ umin when v < umin. the actuator input when umin v umax. respectively. v A ¼ 0.6 3.4 4. actuator limitation. and C ¼ 1.8 4.41.2 4 tc (k) 3. Having covered how the antiwindup compensator can be designed and implemented for a SISO loop. the output of the controller v(k).41 so that it is applicable to many of the control loops described in this book. When the integrator .Internal Process Controls Output and target-point—MAGENTA 4. We describe a general technique by vectorizing Figure 9. and D. let us now focus our attention to the design of a similar compensation strategy for a three-input three-output MIMO level 2 system.

.3146 97.8630 89.2 Figures of Merit Comparison between Three Different Methods (a) Smith Predictor with PI Controller.4999 98.4380 95.0907 97.4776 1 2 3 5 Note: Uses area coverage.6156 92.3143 97.2207 97. and (c) Advanced Nonlinear Controller Obtained for a TC Response with a Step Input of 3%–4% Figure of Merit (%) Percentage of Noise (%) 0 Color Band C M Y K C M Y K C M Y K C M Y K C M Y K State Feedback 97.7087 92. v ¼ umax.3575 92.0717 97.4069 97.5682 97.1406 97.1828 95.4488 97.0907 97.7476 93. The actuator windup occurs when the controller matrix A is not stable and error vector e(k) is nonzero.5516 98.6397 97.3173 96.8273 96.7874 95. time delay actuator saturation.9487 96.7089 98.7065 97.1076 97. (b) State Feedback.4564 97.8153 97.5272 Model Predictive Controller (Not Described) 96.4071 97.7942 96. with a high gain feedback should be able to make the controller matrix stable again.7218 96.3570 96.1762 97.4825 98.6286 97.1433 96.0455 97. The antiwindup compensator.538 Control of Color Imaging Systems: Analysis and Design TABLE 9.9215 97.3993 96.9879 97.1352 97.1949 95.1700 97.3869 97.3468 94.1473 97.1049 97.8857 96.50 Block diagram of the state-space form for a MIMO PI controller.3953 97.1031 95.9666 94.5060 90.4503 Smith Predictor 97. saturates to the maximum allowable actuator limit.9066 86. D x v e B Integrator A C FIGURE 9.0873 97. and sensor noise as various system elements affecting the TC response.1548 96.0096 97.

that is. the TC target needs to be adjusted. Block diagram representation of a MIMO antiwindup compensator with PI When the integrator is not saturated. . L. These adjustments can be provided at a reasonably lower update cycle compared with the basic TC loop described in Section 9. Figure 9.4) for which the closed-loop dynamics are dependent on the eigenvalues of the matrix (A À LC). A practical developability controller operating in Xerox black and white printers through the use of LQR and adaptive parameter estimator is described in Ref.Internal Process Controls 539 D v B–LD e(z) Integrator C umin umax u A–LC – L FIGURE 9. v(k) ¼ u(k). The eigenvalues can be placed at arbitrary locations by selecting the antiwindup gain matrix. This expression represents a combined system of an antiwindup compensator and a PI controller. The compensated controller has new parameter matrices. Another common approach is to augment the level 2 process actuators with a TC target and then develop a four-input three-output control system with low-. which is very similar in structure to a state observer (Section 5. from Equation 9. [16].109 u(k) À Cx(k) À De(k) ¼ 0 (9:110) Multiplying Equation 9.110 by the gain matrix.109 we get a new controller state equation x(k þ 1) ¼ (A À LC)x(k) þ (B À LD)e(k) þ Lu (9:111) Proper choice of matrix L can make the controller matrix (A À LC) stable.11. and high-density DMA measurements.51 shows the block diagram representation of the MIMO PI controller with antiwindup compensator.51 controller. or the machine environment changes. particularly when images with low area coverage are printed. 9.13 OPTIMAL CONTROLS FOR SELECTIVE STATES As the photoconductor and developer materials age. toner age correction is done on both the TC target and the TC sensor. mid-. This is because the desired TC value is adjusted as a function of toner age in order to stabilize the development system. and adding the resulting equation with Equation 9. A-LC and B-LD. In some xerographic printers. if the observability matrix P ¼ [C0 A0 C0 Á Á Á (A0 )nÀ1 C0 ] (with n number of states) is of full rank. L.

1 Note that the above linear representation can also be used for controlling a paper* based density or L* or DEab measurements (extracted with respect to paper) as pointed out in Section 9. respectively. if our goal is to minimize the error vector between the target vector and the measured vector. Let the performance objective be deﬁned in terms of the following function: ~¼1 J 2 NÀ1 P k¼0 a2k ½xT (k)Qx(k) þ uT (k)Ru(k) (9:114) . if U(k) ¼ [Ug(k) Ul(k) Ub(k) UTC(k)]0 is the hard process actuator vector with charge. then for the purpose of developing a four-input three-output stable process control system. a linear state-space model can be written as x(k þ 1) ¼ Ax(k) þ Bu(k) y(k) ¼ Cx(k) where C is the identify matrix y(k) is same as x(k) The system matrix A. then the objective function should be formed with the sums of the squares of the weighted error (or the state) vector. The linear quadratic controller minimizes a selected quadratic objective function. We use the LQR for designing the feedback controller. and x(k) ¼ [Dl(k) Dm(k) Dh(k)]0 is the DMA measurement vector.9.540 Control of Color Imaging Systems: Analysis and Design Ug Ul Ub UTC Charge. xd ¼ DT DT DE is the l m h desired target DMA vector.52. and the TC target [49]. development and TC system at the operating point Dl Dm Dh FIGURE 9. ROS intensity. if. exposure. Â Ã0 Referring to the block diagram of Figure 9. and sensitivity matrix B are expressed as 2 1 0 0 3 2 6 6 B ¼ 6 @Dm 4 @Ug @Dh @Ug @Dl @Ug @Dl @Ul @Dm @Ul @Dh @Ul @Dl @Ub @Dm @Ub @Dh @Ub @Dl @UTC @Dm @UTC @Dh @UTC (9:112) 3 7 7 7 5 (9:113) 6 A ¼ 40 1 0 0 7 0 5. developer bias.52 Four-input three-output MIMO process control system. For example. Another term can be included with the sums of the squares of the actuator values in the objective function to appropriately weigh the desired actuator.

where b ¼ 1=a.117). This will guarantee that the largest eigenvalue will be less than b in magnitude or the transient response will decay faster than bk.3. x we get ~(k þ 1) ¼ aA~(k) þ aB~(k) x x u ~ x(k) þ B~(k) ~u ¼ A~ (9:117) ~ where Ã ¼ aA and B ¼ aB. respectively. Control law: ~ x ~(k) ¼ ÀK(k)~(k) u Gain matrix: Â Ã ~ ~ ~ ~ À1 ~ K(k) ¼ R þ BT P(k þ 1)B BT P(k þ 1)A Recursive equation: Â Ã ~ ~ ~ ~ ~ ~ ~ À1 ~ P(k) ¼ AT P(k þ 1)A À AT P(k þ 1)B R þ BT P(k þ 1)B BT P(k þ 1)A þ Q (9:121) (9:120) (9:119) .Internal Process Controls 541 The scalar parameter a is a positive weight that is chosen by the customer=designer to constrain all the eigenvalues of the overall system to lie inside a circle of radius b 1 in the z-domain.116 and using the deﬁnition of ~(k). since the basic techniques are already described in Section 5. Q and R. the ﬁnal set of recursive algorithms are shown below. We do not show a proof of this method. Optimal gain can be obtained by minimizing ~ subject to J the constraint shown by system model (Equation 9.2. When we increase the weights in Q corresponding to the state x3—the high density state will be emphasized more and affected more (Problem 9. The index N indicates the length of a ﬁnite process.112 in Equation 9. The importance of states and actuators is adjusted by changing the matrices. N can be chosen to be equal to number of cycles over which the optimization function is to be minimized. u Let ~(k) ¼ ak x(k) and ~(k) ¼ ak u(k). Instead. This type of minimization problem will yield the feedback law ~ x ~(k) ¼ ÀK(k)~(k) u (9:118) This can be solved using the standard optimization method involving a Lagrange multiplier.14). Then the new performance index becomes x ~¼1 J 2 NÀ1 P k¼0 ½~T (k)Q~(k) þ ~T (k)R~(k) x u u x (9:115) (9:116) ~(k þ 1) ¼ akþ1 x(k þ 1) x Substituting Equation 9.

Selection of weight matrices: One way of selecting the weights in the quadratic performance equation is shown below. and x3 from their nominal values. then P and K can be determined by solving the following algebraic equations: Gain matrix: Â Ã ~ ~ ~ À1 ~ ~ K ¼ R þ BT PB BT PA Recursive Equation: Â Ã ~ ~ ~ ~ ~ ~ À1 ~ ~ P ¼ AT PA À AT PB R þ BT PB BT PA þ Q Control law: ~x ~(k) ¼ ÀK~(k) u (9:125) (9:124) (9:123) From the above equations. If the process is an inﬁnite stage process (N ! 1). a close form solution is not feasible.123 can be solved to obtain a steady-state gain matrix K. by marching backward in cycles from k ¼ N to 0. a2. Let a1.124.121. The higher the weights. Equation 9. the tighter the tolerance on the control variables. Then the steady-state solution is P(0). Usually. Assume matrix Q to be diagonal (which is more realistic). Therefore. instead of solving an algebraic equation like Equation 9. x2. and a3 be the maximum percentage change deviations of the states x1. This means that Q11 x2 ¼ Q22 x2 ¼ Q33 x2 1 2 3 or Q11 a2 b2 ¼ Q22 a2 b2 ¼ Q33 a2 b2 ¼ b2 1 2 3 (9:128) (9:127) . choose Q such that the overall cost is uniformly distributed between the three states. Equation 9. Once P is found. the gains are clearly dependent on the weights. the feedback gain K attains a constant value.124 is nonlinear. then the total cost is xT Qx ¼ Q11 x2 þ Q22 x2 þ Q33 x2 1 2 3 (9:126) Now. the feedback gain K(k) is time-varying. in this case. If the steady-state form of matrix P(k) is deﬁned as P and the steady-state form of gain matrix K(k) is deﬁned as K. we solve the recursive equation.542 Control of Color Imaging Systems: Analysis and Design Boundary condition: P(N) ¼ 0 (9:122) Steady-state solution: If the control process is ﬁnite (N < 1). Equation 9.

under TRC-based control methodologies (e. Whereas.121. a2 1 Q22 ¼ 1 .121. there is a need to determine the optimum minimal set of gray levels or colors for use during the calibration process. R22 ¼ 1 u2 max 2 . as well as the spatial extent of the test patches when the test patches are rendered in the IDZ areas. halftones.g. particularly when the method corrects time-dependent print quality errors in production color printers. Equation 9. a2 2 Q33 ¼ 1 a2 3 (9:129) Similarly. R44 ¼ 1 u2 max 4 Recursive algorithm (steady-state): If N is chosen to be sufﬁciently large. R33 ¼ 1 u2 max 3 . Since the algebraic equations are nonlinear. the spatial TRCs. although serving a similar purpose. The amount of unnecessary measurement time spent by the printer and the sensing system can be otherwise utilized for making prints useful to the customer. In such systems. and memory requirements of the system. This constant gain can be obtained by either solving the algebraic Equations 9. we choose matrix R to be diagonal. Once P is found.). 9.. media. In particular. A problem usually encountered in such a control system is the ability to deﬁne a minimal sampling procedure for creating and updating the compensation TRCs. it would be computationally more efﬁcient and easier to solve the recursive Equations 9.123 can be used to compute the gain matrix K. etc.124. This results in a reduction in cost and enables more frequent=efﬁcient sampling. We start with the boundary condition P(N) ¼ 0 and solve Equation 9.120 and 9. are generally used to adjust the pixel values (0–255) at a relatively ﬁne resolution to compensate for the characteristics of the print engine so that images are rendered uniformly [50–52]. colors.121 backward from P(N) to P(0). Deﬁning a minimal set of samples becomes increasingly important when the images have to be controlled under multitude of variables.Internal Process Controls 543 Thus Q11 ¼ 1 . with diagonal elements chosen as R11 ¼ 1 u2 max 1 . . TRCs in level 3 control are used for linearizing each separation. time. the feedback gain matrix becomes a constant matrix.123 and 9. particularly for production printers. toner=ink usage. printers.120 and 9. Minimizing the number of samples reduces the computation. The steady-state solution for P is the initial value P(0). or the recursive equations given by Equations 9. one may select an appropriate compensating TRC for a pixel location in a rendered image space based on the contone value of the input pixel.14 OPTIMAL MEASUREMENTS TRCs are updated over time by periodically printing test patches at various gray levels. Printed test patches are sensed by the sensor to determine the appropriate compensation so that the new TRCs provide compensation for the current state of the print engine.

544 Control of Color Imaging Systems: Analysis and Design In this section. The method uses techniques outlined in Ref. a theoretically optimal (or minimal) test patch selection algorithm is developed. in which a measured TRC is approximated in terms of a linear combination of basis functions obtained from experimental data. For simplicity. . whose kth entry is a(k). [53].130 can also be written in matrix form as o ¼ Va þ n (9:131) where V is an eigenvector matrix of size G Â K whose kth column is v(k). or obtained from machine population in the ﬁeld. It is also assumed that the output drift in the system can be modeled as the weighted sum of K eigenvectors. o¼ X k a(k)v(k) þ n (9:130) where o is the above-mentioned output drift vector with G entries G being the number of gray levels v(k) is the kth eigenvector (G Â 1) a(k) is the kth model parameter or weight n is the error vector with independent identical distribution Equation 9. It can be used to model the imaging system with respect to the actuators required for producing compensating TRCs. let us assume that the test patches are gray levels between 0 and 255 digital count. Speciﬁcally. and a is K Â 1. The parameter vector a can be estimated by the least square method as À ÁÀ1 a* ¼ H T H H T y (9:132) where superscript T denotes matrix transpose H is an S Â K matrix S is the number of sample patches used in estimation y is the data vector (S Â 1) that is obtained from experimental data during the runtime control=calibration process The rows of H are selected from the V matrix rows. if the ith patch has a gray level of g(i). The basis vectors which describe the characteristics of the printer are used to determine the optimum minimal set of gray levels or colors for use in calibration or TRC control process [54]. the ith row of H is then a replica of the g(i) th row of matrix V. The eigenvector matrix V is constructed via singular value decomposition (SVD) (Section 3.10) on numerous TRC data samples obtained from a series of controlled experiments in the laboratory.

. if vkj ¼ avki for k ¼ 1. estimation a* is unbiased. and . . The ﬁrst two of these factors are not related to the gray-level sample selection and. if two columns are proportional. (2) the low rank approximation if not all of the eigenvectors are used. . only the estimation error of a is important. then the problem of gray-level patch selection for measurement purpose can be formulated as a method to ﬁnd the matrix H (or equivalently to ﬁnd g(i) for i ¼ 1. In matrix V.Internal Process Controls 545 If the mean-square error is deﬁned as h i E ¼ E ðo* À oÞT ðo* À oÞ (9:133) where o* is the estimation of o and E[Á] stands for the expectation operator. Note that V is orthonormal. and (3) the estimation error of a. S) that yields the least error E. . the potential candidates are ﬁrst pruned. There are three factors that contribute to the error E: (1) the noise n. The set of gray levels used for measurement that possess the maximum dispersion provides the minimum error E which is equivalent to minimizing the trace of a matrix expressed as n Â ÃÀ1 o Optimal gray levels ¼ S ¼ min tr H T H S (9:134) For example. 2. and the expected error energy can be evaluated as Â Ã Â Ã Â Ã X Var½a*(k) E eT e ¼ E DaT V T VDa ¼ E DaT Da ¼ k (9:136) The covariance matrix of a*. K. therefore. it can be evaluated as e ¼ VDa (9:135) where Da ¼ a* À a is a K Â 1 estimation error vector. . . if the G Â 1 vector e is deﬁned as the error introduced by estimation inaccuracy. speciﬁcally. . as described in Ref. yielding X k (9:137) hÀ ÁÀ1 i Var½a*(k) ¼ tr½Covða*Þ ¼ tr c H T H (9:138) To reduce the computation time. . [54] can be evaluated as À ÁÀ1 Covða*Þ ¼ c H T H where c is a constant. 2.

perform steps c and d below. where R is the number of the rows in V that have survived pruning. For i ¼ 1 to S: Calculate the change of tr[(HTH)À1] with the replacement of g(i) with g(i À 1). then À ÁÀ1 U À1 ¼ H T H À h(i)h(i)T þ h(j)h(j)T (9:139) where h(i) is the ith column of V. a brute force search may ﬁnd the sample gray levels that minimize tr[(HTH)À1]. This is due to the fact that if column j were selected. only one vector in H is replaced. because inversion is computationally expensive for large matrices. an iterative method such as Newton–Raphson or a genetic method may be applied. Equation 9.h À ÁÀ1 i ¼ H T H h(i)h(i)T H T H 1 À h(i)T H T H h(i) (9:140) (9:141) . Using the Woodbury identity twice. However. b. . column j will be eliminated from the selection process. we can take advantage of the fact that. column j can be removed from the matrix. and the calculation can therefore be simpliﬁed. The calculation is to compute tr[(HTH)À1] RS number of times (assuming two or more samples may share the same gray level). Otherwise. Calculate the change of tr[(HTH)À1] with the replacement of g(i) with g(i þ 1). From the calculation in step c.139 can be simpliﬁed as Ã Â ÃÂ U À1 ¼ W À1 À W À1 h(j)h(j)T W À1 1 þ h(j)T W À1 h(j) where À ÁÀ1 W 1 ¼ H T H À h(i)h(i)T hÀ ÁÀ1 À ÁÀ1 i. on each pass. When S is larger. The calculation in step c can be substantial. . If U denotes the updated matrix of HTH after g(i) is replaced by g(j). Therefore. c. d. . Select an initial set of gray levels g(1).546 Control of Color Imaging Systems: Analysis and Design the proportional factor a is less than 1. g(S). The iteration ends when no replacement can make any improvement. One simple method is as follows: a. For each iteration. When S is small. ﬁnd the one that minimizes tr[(HTH)À1]. replacing it with column i will always reduce tr[(HTH)À1]. . update the H matrix based on another set of S gray levels and repeat step c until there are no more remaining sets of S gray levels to select. especially when K (model rank) is large. g(2).

Internal Process Controls 547 Since (HTH)À1 is known from the results of the ﬁnal step in the iteration.134 for spatial TRC data. These measurements serve as a reference for comparison of TRCs derived from subsets of gray levels. The gray levels used in the procedure are provided in Table 9. Prior experience shows that this number is adequate when signiﬁcant drift in the print engine response is not present. Also. and the selection process may be conducted in a hierarchical manner to further reduce computation. The error is calculated by ﬁrst deriving all gray levels (32) from a given subset through the use of basis vectors. A system is assumed with a small number of gray levels (32. TRCs are generated using the optimal levels derived above. These ﬁrst printing and measurement results are used to 1. the above calculation essentially contains only two matrix–vector multiplications. The derived 32-level TRCs are then . Derived TRCs are compared to the fully measured TRCs. Derive the optimal gray levels for future calibration updates. More basis vectors can be included depending on the level of accuracy needed.13 As an example. 4. A second printing of the 32 levels is performed and used for the following steps: 1. 4. namely 2. In many typical control=calibration applications. Optimal gray levels are derived for three separate cases: 2 levels. the eigenvectors are smooth functions. No matrix inversion is performed. rather than 256) to make the demonstration tractable with a reasonable amount of effort and a small number of test prints and measurements. and 8 levels.3 below. 4 levels. we ﬁnd the optimal gray levels by solving Equation 9. The solution for the reduced resolution is used as the initial condition to run the same procedural steps a through d again for the full resolution calculation. In our example. the optimization procedure ﬁrst executes the above-described procedural steps a through d with a reduced resolution. a reduced set of two basis vectors is selected as the system characterization. and 8 levels. Example 9. 2. Speciﬁcally. 2. Complete print engine state (spatial TRCs) is measured for t > 0. TRCs are generated using several equispaced subsets of gray levels. intuition can be used in selecting combinations of gray levels so that the number of combinatorial searches required can be minimized. The quality of TRCs derived from equispaced levels is compared with the quality derived from levels determined by using the optimal gray level method. 3. A ﬁrst printing is performed using all of the 32 gray levels in a calibration target where a gray strip for each level spans across the page (Figure 9. Measurements are taken of the spatial nonuniformity using a scanner for each gray level to obtain spatial TRC data.53a). Derive TRC eigenvector matrix that characterize the system.

169. 161. which can be performed in the inter document zone FIGURE 9. TABLE 9.54.53 Photoreceptor showing calibration strips for on-belt sensing for calibrating streak compensating TRCs. 241 Equispaced Values 65.3 Optimal and Equispaced Gray Levels (0–255) Number of Patches 2 4 8 Optimal Values 137. 225. In Figure 9. 177. 137. 233 137. 81. For each subset. 233 113.548 Photoreceptor Control of Color Imaging Systems: Analysis and Design ……… …… … Customer image area (a) Non optimal gray levels requiring printing in the customer image area Photoreceptor Customer image area Customer image area Customer image area (b) Optimal gray levels requiring only minimal printing. the mean-square error (MSE) over the 32 levels is plotted. 169. 153. 145. . 97. 113. 233. 193 33. 4-. and 8-level subsets is calculated across the spatial dimension for all gray levels. 49. 225 17. 209. 201. 225. The resulting data points are plotted and it can be observed that the optimal-set error incurred by using the optimal gray levels is roughly half of the equispaced-set error incurred by using the equispaced levels. the MSE for each of the 2-. 241 compared to the measured 32 levels.

2 0 Equal-space set 549 Mean square error 2 3 6 4 5 Number of gray level strips 7 8 FIGURE 9. say b22. Show the evolution of the states for xd ¼ [À550 À100]T and x(0) ¼ [À588 À87]T with respect to 10 pitch numbers.6 0. Calculate the steady-state values of the actuator vector. Show that the steady-state error is held to zero. b.1 Let the state-space equations for the level 1 electrostatic system be ! ! 0:9798 0 x(k þ 1) ¼ x(k) þ v(k) 0 1 0:4574 14:9315 ! 1 0 x(k) y(k) ¼ 0 1 1 0 Use the control law shown in Equation 9. À0:025 0:0469 a.4 0.2 1 0. c.1. Mean-square error from the two gray-level selection methods for 2. and 8 PROBLEMS 9. 4. for (a) and (b). even when one of the elements of the Jacobian matrix.Internal Process Controls Mean square error between estimated and measured TRCs (w/testing data) 1. What are the eigenvalues of the open-loop system with and without the integrator in the loop? Is the system open loop stable in both . v(k).54 patches.2 Determine the controllability and observability matrices and then ranks for Problem 9.8 Optimal set 0. 9.26 with the feedback gain matrix given ! 0:8165 0 as K ¼ . is lowered by 20%.

Find the expression for the steady-state actuator U. 9. If all the elements of matrix B are increased by 20% and the gain matrix. e. g. have changed by 20%. Assume the gain matrix as K ¼ BÀ1 when the integrator is replaced with the unit gain. It is now required to design a three-input three-output developability control loop using three process actuators. Ul(k) ¼ X(k). and Dh(k) represent the three different DMA measurements shown in the state vector. indicated by the parameter. Ub(k) ¼ Vbias(k). measured every pitch. Determine the open and closedloop transfer functions for the new structure. The charging system is characterized by the Jacobian matrix B. Comment on the stability. Ubo ¼ À200 V} as shown below: 1 0 6 x(k þ 1) ¼ 4 0 1 1 0 6 y(k) ¼ 4 0 1 0 0 2 0 0 2 3 2 0 0:06813 7 À3 6 0 5x(k) þ 10 4 0:27346 1 0 7 0 5x(k) 1 3 1:55592 À0:06047 À0:23242 À1:43732 3 5:69587 7 21:52804 5v(k) 94:84767 . Write the open. but the voltages at the nominal operating point. Let Dl(k).4 Consider the development control system of Example 9.3 Consider the electrostatic control system of Figure 9. and measurements from the DMA sensor. Which controller gives the lowest steadystate error? f. k. and the development bias.4. b. d. Is the steady-state actuator vector dependent on the gain matrix? Explain the reasons for your answer. determine the percentage change in steady-state errors with and without integrators in the loop. 9. Determine the steady-state error between the desired exposed and unexposed voltages to the measured voltages. 9. at the nominal operating point. Use any convenient signals as inputs and outputs. Ulo ¼ 4 ergs=cm2.and closed-loop transfer functions of the system. the grid voltage.550 Control of Color Imaging Systems: Analysis and Design cases? Close the loop with the gain matrix of Problem 9. Dm(k). determine the percentage change in steady-state errors with and without the integrators. at the nominal operating point. x(k). is equal to that used in d. a. c. The development system is characterized by the Jacobian matrix B.1. Determine the steady-state error.5. If B matrix has not changed. Determine the eigenvalues. K. Vho and Vlo. The linear state variable description of the control system is characterized by the Jacobian matrix at the nominal operating point {Ugo ¼ À600 V. Ug(k) ¼ Vg(k). Replace the integrator with the unity gain. the exposure intensity. Compare the results of b with d.5 The cost of an electrostatic sensing system can be reduced by removing the charge sensor (ESV) and associated level 1 control loop.

Use identity reference TRC and direct inversion process. Process magenta separated image through the inverted TRC and printer model. Compare the results of step b by running the control simulation with the charge. x. a.92 242. the vector.4. TABLE 9. 0. 9.00 230. Find the gain matrix for placing the poles at location [0. In the third column.00 255.00 51.00 77.31 203.2. y(k) ¼ x(k).8).00 8. x(k) ¼ 4 Dm (k) 5.00 102.7 Let the vector U0 contain the input patch gray levels (or area coverages) for * controlling DEab from paper for magenta separation shown in Table 9. d.00 26. c.20 43.89 255. 0.00 153. design the antiwindup compensator gain matrix and show the performance with and without the compensator. 9. v(k) ¼ 4 DUl (k) 5.6 For a level 2 controller (Figure 9.00 2 .00 * DEab with Respect to Paper 0.00 179. the measurement-actuation updates are executed at every pitch. Show the time evolution of the states for a step response with respect to pitch number using recursive solution of Equation 4. b.2] using Equation 9.00 204.61 161. development models from Chapter 10.37 26.76 128. They are shown in the second column.Internal Process Controls 551 3 2 3 DUg (k) Dl (k) where.2.77 88. contain the * normalized values transformed to gray levels of the DEab from paper measurements are shown.118.45. The controller uses a Dh (k) DUb (k) gain matrix and the integrator modeled by the following Equation: e(k þ 1) ¼ e(k) À Bv(k) v(k) ¼ þKe(k) where e(k) ¼ xd À x(k) and xd is the desired state vector. Are there other reasons why this type of control approaches are more suitable than level 1 and 2 architecture? Comment.73 75.00 128.4 * DEab from Paper (Normalized) with Respect to Gray Levels Number of Patches 0 1 2 3 4 5 6 7 8 9 10 U0 0.

simulate the situation with integrator windup. US Patent 5.10 using the LQR method. 3. in Proceedings of the IEEE International Conference on Control Applications. p2 ¼ 0. 28.189.255. and T. May 19.185. p3 ¼ 0 b. 5. Henderson.7. 5. Italy. . 5. UC Berkeley.471. M.A.e.K. A multilevel modular control architecture for image reproduction.E. REFERENCES 1.5 Give a step change in DMA targets. Comment on the results.E. 2. Show the equation for the steadystate value and plot it as a function of weights used in TC target. 2004.673. Salt Lake City. UT. p2 ¼ À0. Compare the controllers. 1998. Apply the Smith predictor and a PI controller and simulate the performance again.918. Mestha and P.9 9.K. Padmanabhan.10 when the TC system has a ﬁxed transport lag of 3 and 10 cycles.13 Describe the relationship between TC control and a typical inventory management system. Dh. Wang. 5. NIP20: Proceedings of IS&T’s International Congress on Digital Printing Technologies. Electrostatic control with compensation for coupling effects.A.. 1–4. Other US Patents 5.940. T. Butler. p3 ¼ À0. Thieret. Consider a scenario where the level 1 loop of Example 9. 1998.441. 578–585. 5. Varaiya.14 For a four-input three-output controller with TC target as one of the actuator. Process magenta separated image through the inverted TRC and printer model. 9.754.R. Sketch the expected convergence plots for case a and b for state variables {Vh. L.552 Control of Color Imaging Systems: Analysis and Design 9. Thieret.345.462. Sep. ~ ~ 9. with all other weights unchanged.5. Design a antiwindup compensator with high gain L and simulate the integrator recovery and the TC performance with antiwindup compensator.087.12 Simulate the TC performance for a system with the state feedback controller and estimator in Example 9.461. 5. Mestha. T. US Patent 5. and M. Comment on the results.Vl.231. Dl} at every print for a total of 30 prints. Nov. Dm. 5. L. Control advances in production printing and publishing systems. 31–Nov. Towards a layered view of control. Y.085. 5. 9.A. 1995.8 Obtain an inversion using control-based inversion for Problem 9. pp. Trieste. 4. closed-loop eigenvalues ¼ 0) and level 2 loop has the following closed-loop poles: a.172. Mestha. 9.452. 7. CA 94720. L.162.6 is closed for every 5 prints. Use identity as the reference TRC.11 For TC system with PI controller umin ¼ 0 and umax ¼ a reasonable limit.1 is closed for every print and the level 2 loop of Example 9. P.315.10 Design the gain matrix for the PI controller of Example 9. Oct. Compare your TC performance with the pole-placement technique. Let level 1 loop be designed for a dead beat response (i.313. design the LQR to emphasize the TC target.K. 9. p1 ¼ À0. p1 ¼ 0.5. Scheuer. Method and control system architecture for controlling tone reproduction in printing device.

227.652. 7.115. 1998. Robust eigenstructure assignment in state feedback control. Charge control system.749. Optimal time-sequential sampling of tone reproduction function. US Patent 5. in Proceedings of IEEE Conference on Decision and Control. Australia. Other US Patents 5.191. 1998.P. Method to model a xerographic system.258. Method for measurement of tone reproduction curve using a single structured patch.967.436. S. 4.382.205.928.337.162. Other US Patents 4. 7. 5.Y. 4.395. Prentice Hall.383. 7.840.391. ISBN 1-880156-03-2. 5.556.792. MN.253.816.543. 1985.109.978. 6. 4. S. US Patent 5. 6. 7. J. NJ. 5. 6. Sep.267.705.576. 5. M.285. L.152.K. 4. L. K.811.. 6. Khargonekar. 5. T. 6. Launsby. 7.043.564.697.243. Understanding Industrial Designed Experiments. T. Other US Patents 5.896. 5. Jul.R.117. Developed mass per unit area (DMA) controller to correct for development errors.045. 4.P.033.239. US patent 4.211.A. Colorado Springs.811. 5. Other US Patents 4.870.564.R. Hayes. 5.341.P.266. 1995.023. Nov.946.853. Mestha.K. T. Sep. T. SA 5042. 12. 7. 9.122. 6.203.918. III and J. K. 12.816.035.812. 6.456.796. 6. Koditschek. 611.205.783. S. 5. US Patent 5. 16. 5.831.217. 6. Raj. Numerical Analysis Report NA=2=83. 4. Minneapolis. 1980.866. 7. 17. Hubble. Feb. Image forming apparatus with predictive electrostatic process control system.. 1982. May 5. Adaptive process controller for electrophotographic printing.665. Maebashi.741.519. 16.262. and K. Englewood Cliffs. Nakai and Y.426.018. and D. 23. 13.060.804. 14. Other US Patents 6.523.819. 21. Bedford Park.989. Mar.113.200. 4. AIR Academy Press & Associates. 1984. 6. 5.212. Jun. Development control of a reproduction machine. Kobayashi.256.113. Japan. 10. 1996.270. US Patent 4. US Patent 4. Ono.568. Jun. High speed electrostatic voltmeter. System for identifying areas in pre-existing image data as test patches for print quality.F.869. Other US Patents 4. 5. 20. T. Ogata.907. May 27. US Patent 4.835.997.187. Ono.K. Y. F. 2nd edn. 7. 6. 13. Image recording apparatus for controlling image in high quality and image quality control method. Henderson. E. 4. 4.331.204. Mestha.842. 4. 6. Toward a control oriented method of xerographic marking engine. 6. 11.E.246.010. 5. 2003. 6.749.122. 6. 1993.749.126. Algorithm coded in Control System Toolbox for use with MATLAB. NJ. 5. Kobe.800. 4. Sato. K.968.461.639.024. Kautzky and N. 5.003. US Patent 5.416.204. 6. L. 2005. 14–16.348. 5. Dec 11–13. Linear Systems.198.383.401. S. Flinders U.330. Other US Patents 3.832. 12.985. 5.922.741. 5. US Patent 5. 5. 5.371.633. T.099.639.323. 6. 18.A.352.243. Sim and P.703.054.167.553. 4.034.A.B.013.005. Jr.481.634. 4. 7. Prentice Hall. 6. Upper Saddle River. 19.972. 2006. 26.538. Sep.315.173. 5. Infrared reﬂectance densitometer. Williams. 4. 4. 22.582. 19.886. Other US Patents 4. 1995. L. 27.065. Dianat. 4.873. 10.750. US Patent 6.450. Fantozzi.370. G.200.660. 25.169. Image recording apparatus for controlling image in high quality and image quality control method thereof.250. 8.270.221. 1995.544. P. 1995.481.497. Schmidt and R.T. 5. 5. 4.650. Sato. US Patent 5. Parisi.371. Other US Patents 4. Jackson.220.033. Discrete-Time Control Systems. The Math Works. and K. 4. Image forming apparatus with transfer voltage control for transferring toner patterns.606.903. 7.162.005. Other US Patents 7. CO. 1996.021.433.816. 1995.387. 15. L. Martin. Jul.684. B.989. in Proceedings of the 2006 American Control Conference.912.771. 1992. 5.166. Kobayashi.G.451. Kailath.868.383.866. MA.391.630.717. .988.229.Internal Process Controls 553 6. School of Mathematical Sciences.853. 1980.550.705.965. Thieret.481.477. 6.553. 4. 5.436.K. Li.J.741.257.384.320.639. Mestha.694.165.122.986. Nichols.057. May 13. Nov. US Patent 5. 4. Mestha. US Patent 5.576.853. 4..243. Wang.882. 5.K.021.

M. J. 30.298.861. Marking engine and method to optimize tone levels in a digital output system.727. Sim. J. Hamby. MD.845. Anti-windup proportional plus integral controller. 2005. US Patent 5.133. G. 663–666. 5. Apr. 1977.572. 1992.C. New York.698. F. 1994. Roy.J. DeBoer and D. ISBN 0-13-040361-X.J.L. 3. Toner age calculation in print engine diagnostics. 5.K. E.298.678. Friedland. 8.974.939. Jr. 36. Filter for reducing effect of noise in TC control. Other US Patents 3.336.W. 6.646. 5. P. Jul. Sep. 16. Grace. Englewood Cliffs.036. 34. and J.025. 6.W. NJ. Love. P. Feedback toner concentration control for an imaging system. and A.353.733.032.R. 4. 1(2–6).R. Nortcliffe and J.032.N.141. 9. 5. Signature sensing for optimum toner control with donor roll. Scheuer. Quantiﬁcation of toner aging in two component development systems. 5. 1999. and E.192. Pacer.A. Automatic compensation for toner concentration drift due to developer aging. Maltz. . ISBN 0-13-212043-7.221. Jan. 7.175.980. 30th Annual Conference of IEEE. E.214.121. 9.M. Jr.909. W. 6. Crean and M. 5.604. C.227.A. Jr.P. 5.147. Donaldson. and J.369. D. 28.J.397.106. 31.830.S.010.S. B..554 Control of Color Imaging Systems: Analysis and Design 24. 32. Prentice Hall.829. 38. C.045. 6.971.012. ISBN 0-471-57795-2. G. 28.901. Hamby.213. 29. P. R. US Patent 4.643.Y.679.498. Mestha and S. Ward. B. 2001. Padmanabhan.W.103. 39. 40. Nov. Nov. Dianat. TRC smoothing algorithm to improve image contours in 1-D color controls. 3.710. Balakrishnan and G. 2001.581. Toner concentration sensing apparatus. Other US Patents 4.. Toner concentration control for an imaging system.892.167.326. M.115.419.286. Park. New York. MacDonald. Y. 6. P. 25. US Patent 5.958. D.903.401. L. State Variables for Engineers. 5. Van Cleave. Donaldson. J. Smith predictor control for water pressure control system with time delay.W. 29.K. 43(1).J. P. Gross.S. Birkhauser. 2004. US Patent 6. US Patent 5. 24. John Wiley & Sons. US Patent 6. 1995.729.W. Donaldson. Borton. pp.J. US Patent 5.W. Mestha.144. Scheuer.. 1998. 4. Wang and L. Lee. Applied Optimal Control & Estimation: Digital Design & Implementation.756. ISA Transactions.Y. 2nd edn. Buranicz. 2003.514. MA.142. Nov.388.M. 6. 35.J. Buranicz.812. Jan.194. US Patent 5. Smith. 6. DeRusso.032. P. Englewood Cliffs. Desrochers.895. E.A. ISBN 0-07-286376-5. Digital Control System Analysis and Design.L. Gros