Muhammad Rzi Abbas Department of Mechatronics and Control Engineering
muhammadrziabbas@uet.edu.pk Lecturer, Mechatronics Dept. University of Engineering and Technology, Lahore • Decision Theoretic approaches may come handy if we want to find objects within the image and identify, or classify, those objects. • The central assumption in this approach is that the image depicts one or more objects and that each object belongs to one of several distinct and exclusive pre-determined classes, i.e. we know what objects exist and an object can only have one particular type or label. • Three components of such pattern recognition process are: • An object isolation module • A feature extraction module • A Classification module • Output of one module is the input of the next module that follows. • Object Isolation Module operates on a digital image and produces a representation of the object. • Feature Extraction Module then abstracts one or more characteristic features and produces a feature vector. • Classification Module then uses this feature vector to identify and label each object. Object Isolation Module • Nothing but segmentation • We have studied segmentation in detail • Once segmentation is done we have identified the objects which we wish to classify or recognize. Feature Extraction Module • Now the task is the extraction of features which are characteristic of the object and which will be used in the Classification Module. • Selection of feature is an extremely important task, since all subsequent decisions will be based on them • Normally number of reasonable feasible potential features are identified, tested and finalized to be used in actual application Feature Extraction Module • While selecting a feature following things should be kept in mind • Independent (a change in one feature should not change the value of another feature significantly) • Discriminatory (each feature should have a significantly different value for each different object) • Reliable (feature should have the same value for all objects in the same class/group) Feature Extraction Module • It is now evident that computational complexity of the pattern recognition exercise increases rapidly as the number of features increase and hence it is desirable to use the fewest number of features possible, while ensuring a minimal number of errors. Simple Features • Most features are based on the size of the object or on its shape. • One obvious size-based feature is area • This is simply the number of pixels comprising the object multiplied by the area of a single pixel. • A modified version of area is ‘Integrated Optical Density’ (IOD). • It is equivalent to the area multiplied by the average gray-level of the object and essentially provides a measure of the ‘weight’ of the object. Simple Features • The length and the width of an object also describe its size. • But if such feature is to be used then one should keep in mind that they are not orientation independent. • The minimum bounding rectangle is a feature which is related to the idea of length and width. • This is the smallest rectangle that can completely enclose the object. Simple Features • The distance around the perimeter of the object can be useful for discriminating between two objects. • Depending on how the object is represented, and this in turn depends on the type of segmentation used, it can be quite easy to compute the length of the perimeter which makes it an attractive feature for industrial vision applications. Simple Features • Features which encode the shape of an object are usually very useful for the purpose of classification. • Two popular features are Rectangularity and Circularity. • Rectangularity may be measured through following popular measures: • Ratio of the area of the object to the area of the minimum bounding rectangle. Simple Features • Rectangularity may be measured through following popular measures: • Ratio of the area of the object to the area of the minimum bounding rectangle.
• This feature takes on a maximum value of 1 for a perfect rectangular shape
and tends toward zero for thin curvy objects. Simple Features • Rectangularity may be measured through following popular measures: • The second measure is the aspect ratio and is simply the ratio of the width of the minimum bounding rectangle to its length. Simple Features • The most commonly used circularity measure is the ratio of the area to the square of the perimeter length. • This assumes a maximum value for discs and tends towards zero for irregular shapes with ragged boundaries. • Machine Vision by David Vernon, Published in 1991 • Chapter 6 (Section 6.3)