You are on page 1of 13

10/8/2015

Digital Signal Processing

10/8/2015 1

• References

• Digital Signal Processing, 4/E


John G. Proakis, Northeastern University
Dimitris K Manolakis, Massachusetts Institute of Technology,
Lincoln Laboratory

• Digital Signal Processing: A Computer-Based Approach, 4/e


Sanjit K. Mitra, University of California - Santa Barbara

• Digital Signal Processing Using MATLAB 3rd Edition


by Vinay K. Ingle, John G. Proakis

10/8/2015 2

1
10/8/2015

Tentative course outline

• Discrete Fourier Transform

• Digital Filters

• Multirate signal processing

10/8/2015 3

frequency domain signal processing

10/8/2015 4

2
10/8/2015

Summary of Fourier analysis

10/8/2015 5

Summary of Fourier Analysis

• Fourier series for continuous time periodic signals

• Fourier transform for continuous time aperiodic


signals

• Discrete-Time Fourier Transform (DTFT) for discrete


time aperiodic signals

• Discrete Fourier Transform (DFT) for discrete time


periodic signals
10/8/2015 6

3
10/8/2015

Discrete Time Fourier Transform (DTFT)

• The kernel of the transform is continuous function of the


frequency


• Sufficient condition for convergence 𝑛=−∞ 𝑥,𝑛- < +∞

• 𝑋(𝜔) is continuous and periodic with period 2𝜋


10/8/2015 7

dot product and the projection concept

0
𝑒2 =
1

4
𝑥=
3

1
𝑒1 =
0

10/8/2015 8

4
10/8/2015

• example

• The DTFT is real for an even sequence


10/8/2015 9

Symmetry properties of the DTFT

10/8/2015 10

5
10/8/2015

• example
Determine and sketch 𝑋𝑅 𝜔 , 𝑋𝐼 𝜔 , 𝑋(𝜔) , 𝑎𝑛𝑑  𝑋(𝜔) for

𝑥 𝑛 = 𝑎𝑛 𝑢 𝑛 , −1 < 𝑎 < 1

10/8/2015 11

Gibbs phenomenon

Partial sum:

we will recall this later!!


“Windowing effect”

10/8/2015 12

6
10/8/2015

• Ringing artifacts in coded images “Gibbs phenomenon”

• The loss of high frequency near sharp edges

• Common in:
– Compression
– Band-limited transmission

10/8/2015 13

• Pre-echo in audio compression “Gibbs phenomenon”

• The sound is heard before it occurs.


Coded Castanet Signal

Original Castanet Signal

10/8/2015 14

7
10/8/2015

Energy Density Spectrum of Aperiodic Signal


“Parseval relation”

10/8/2015 15

• The spectrum 𝑋(𝜔) is a complex valued function, in general

• The energy density spectrum of 𝑥(𝑛) is 𝑆𝑥𝑥 (𝜔)

does not contain any phase information !!

10/8/2015 16

8
10/8/2015

example
• Determine and sketch the energy density spectrum 𝑆𝑥𝑥 (𝜔) for the sequence
𝑥 𝑛 = 𝑎𝑛 𝑢 𝑛 −1<𝑎 <1

10/8/2015 17

example

phase
information
is lost in the
spectrum

10/8/2015 18

9
10/8/2015

magnitude vs. phase in image reconstruction

Original image Energy spectrum

10/8/2015 19

magnitude vs. phase in image reconstruction

Reconstruction from magnitude only Reconstruction from phase only

location information is preserved in the phase!!


10/8/2015 20

10
10/8/2015

Properties of the DTFT

10/8/2015 21

The relation between the energy density spectrum and the


autocorrelation of a sequence
“Wiener-Khintchine theorem”
Recall that the autocorrelation 𝑟𝑥𝑥 𝑙 of a sequence 𝑥 𝑛 is
defined as
𝑟𝑥𝑥 𝑙 = 𝑥,𝑙- ⊛ 𝑥,−𝑙-
Taking the DTFT

2
𝐷𝑇𝐹𝑇 𝑥 𝑙 ⊛ 𝑥 −𝑙 = 𝑋 𝜔 𝑋 −𝜔 = 𝑋(𝜔)
= 𝑆𝑥𝑥 (𝜔)
Then

𝑆𝑥𝑥 𝜔 = 𝐷𝑇𝐹𝑇 𝑟𝑥𝑥 𝑙 = 𝑟𝑥𝑥 (𝑙)𝑒 −𝑗𝑙𝜔

• both the energy density spectrum and the autocorrelation


function contain the same information, phaseless.
10/8/2015 22

11
10/8/2015

example
• The power density spectrum of white noise is flat.

10/8/2015 23

Linear convolution using the DTFT

According to the convolution theorem if 𝑦 𝑛 = 𝑥 𝑛 ⊛ ℎ 𝑛

Then the DTFT (𝑦 𝑛 ) is 𝑌 𝑒 𝑗𝜔 = 𝑋 𝑒 𝑗𝜔 𝐻 𝑒 𝑗𝜔

Then 𝑦,𝑛- can be computed as IDTFT(𝑌(𝑒 𝑗𝜔 )

10/8/2015 24

12
10/8/2015

example
• Determine the autocorrelation of the sequence 𝑥 𝑛 = *1,1,1+
𝑟𝑥𝑥 𝑙 = 𝑥 𝑙 ⊛ 𝑥 −𝑙 , 𝑋 𝜔 = 𝑋 −𝜔 = 1 + 2cos(𝜔)
• Then
𝑋 𝜔 𝑋 −𝜔 = 1 + 2cos(𝜔) 2
= 3 + 4 cos 𝜔 + 2 cos 2𝜔
= 3 + 2 𝑒 𝑗𝜔 + 𝑒 −𝑗𝜔 + 𝑒 𝑗2𝜔 + 𝑒 −𝑗2𝜔

Taking the IDTFT, we have 𝑟𝑥𝑥 𝑙 = *1,2,3,2,1+

10/8/2015 25

13

You might also like