EECE\CS 253 Image Processing

Lecture Notes: Sampling Lecture Notes and Aliasing

Richard Alan Peters II

Department of Electrical Engineering and Computer Science
Fall Semester 2011

n i

This work is licensed under the Creative Commons Attribution-Noncommercial 2.5 License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc/2.5/ or send a letter to Creative Commons, 543 Howard Street, 5th Floor, San Francisco, California, 94105, USA.

Warning:

Image Resampling Can Lead to the Jaggies!

The jaggies!
zoomed zoomed2 2

20 October 2011

1999-2011 by Richard Alan Peters II

This is a bad way to do it.

Downsampling (Decimation)
>> >> J J= = I(1:2:R,1:2:C,:); I(1:2:R,1:2:C,:);

E.g.: every 2nd pixel in every 2nd row

20 October 2011 1999-2011 by Richard Alan Peters II 3

Bad, bad, very bad.

Downsampling (Decimation)
>> >> J J= = I(1:2:R,1:2:C,:); I(1:2:R,1:2:C,:);

E.g.: every 2nd pixel in every 2nd row

20 October 2011 1999-2011 by Richard Alan Peters II 4

Recall:

Power Spectrum from Discrete Fourier Transform

DFT

20 October 2011

1999-2011 by Richard Alan Peters II

Recall:

Power Spectrum from Discrete Fourier Transform

decimated image

DFT

The DFT of an image is the same size as the image.

20 October 2011

1999-2011 by Richard Alan Peters II

power spectrum
6

Recall:

The Scaling Property of the FT

If then

F {I}(v, u) = I (r, c) ei 2 ( uc+vr ) dcdr,

r c F I , ab I av , bu . { }( ) a b
This Thisimplies impliesthat thatif ifan animage imageis isreduced reducedin insize, size,its its features featuresin inthe thespatial spatialdomain domainbecome becomesmaller smallerand and its itsfeatures featuresin inthe thefrequency frequencydomain domainbecome becomelarger. larger.

20 October 2011

1999-2011 by Richard Alan Peters II

Recall:

The Uncertainty Relation

space frequency

FT FT

If xy is the extent of the object in space and if xy is its extent in frequency then
x y uv 1 16 2

space

frequency

FT FT

A Asmall smallobject objectin inspace space has hasa alarge largefrequency frequency extent extentand andvice-versa. vice-versa.
1999-2011 by Richard Alan Peters II 8

20 October 2011

Contradiction?

Effect of Decimation on the DFT of an Image

1. 2. 3.

Decimation of an RC image, I, by a factor of n results in an R/n C/n image, J. The DFT of image J is the same size as J. The uncertainty relation implies that the FT of J should be (R-1) (C-1) < n R/n n C/n R C.

Q: How can these 3 facts be true simultaneously? A: The FT of J folds over or aliases itself on the DFT of J because the DFT is defined on a torus.
20 October 2011 1999-2011 by Richard Alan Peters II 9

Discrete FT is on a Torus

Recall: the theDFT DFTof ofaa Recall: digitalimage imageassumes assumes digital theimage imagehas hasaa the toroidaltopology topology toroidal

20 October 2011

1999-2011 by Richard Alan Peters II

10

Discrete Fourier Transform is on a Torus

20 October 2011

1999-2011 by Richard Alan Peters II

which whichimplies impliesthat that the result also has the result also hasaa toroidal toroidaltopology. topology.

11

Point 1 (p. 9)

Discrete FT is on a Torus
Ifthe theimage imageis ismade made If smallerthe thetorus torusis is smaller likewise likewise

20 October 2011

1999-2011 by Richard Alan Peters II

12

Point 2 (p. 9)

Discrete FT is on a Torus
and andthe theDFT DFTis isthe the samesize. size. same

20 October 2011

1999-2011 by Richard Alan Peters II

13

Ideal PS of 2 Decimated Image

zoomed2 2 zoomed

Onemight mightassume assume One thatthe thesmaller smallerDFT DFT that isthe thecentral centralregion region is ofthe thelarger larger but but of itsnot. not. its

20 October 2011

1999-2011 by Richard Alan Peters II

14

Allof ofthe thelarger largerone oneis is All presentin inthe thesmaller smallerone. one. present

To Make Actual PS of 2Decimated Image:

low freqs. high freqs. of cols.

of cols & rows

low freqs. of rows

high freqs. low freqs. of cols. high freqs. of rows of cols & rows

20 October 2011

1999-2011 by Richard Alan Partition intoII thesefour fourparts, parts, Partition {{ II }}Peters into these

15

Actual PS of 2 Dec:

20 October 2011

1999-2011 by Richard Alan Peters II

four fourparts partsthat thatare arecontiguous contiguous on onthe thefull-sized full-sizedDFT DFTtorus. torus.

16

Eachof ofthe the4 4PS PSregions regions Each formsaatorus. torus.The The4 4torii torii forms aresuperpositioned superpositionedonto onto1. 1. are

Actual PS of 2 Decimated Image

low freqs. high freqs. of cols.

of cols & rows

low freqs. of rows

high freqs. low freqs. of cols. high freqs. of rows of cols & rows

20 October 2011

1999-2011 by Richard Alan Peters II

17

DFT Aliasing on the Torus

The4 4superpositioned superpositionedtorii torii The aliasthe thehigh highfrequencies frequencies alias ontothe thelow lowfrequencies. frequencies. onto
20 October 2011 1999-2011 by Richard Alan Peters II 18

PS of Original and PS of 2 Decimated Image

Aliased Aliased frequencyfrequencyplane plane features features

original original
20 October 2011

decimated decimated2, 2,zoomed zoomed2 2

1999-2011 by Richard Alan Peters II 19

lead leadto tojaggies jaggiesin in thedecimated decimatedimage. image. the

Original Image and 2 Decimated Image

original original
20 October 2011

decimated decimated2, 2,zoomed zoomed2 2

1999-2011 by Richard Alan Peters II 20

First Alternative Explanation of Aliasing

The aliasing phenomenon can be described in terms of the so-called wagon wheel effect. The name comes from the appearance in a 24 frame per second movie of a wagon wheel rotating slightly faster than 12 frames per second. It appears to be rotating slowly backward.

20 October 2011

1999-2011 by Richard Alan Peters II

21

Second Alternative Explanation of Aliasing

The aliasing phenomenon can also be described in terms of the convolution property of the Fourier Transform. In fact, this is the way it is usually explained.

20 October 2011

1999-2011 by Richard Alan Peters II

22

The Sampling Function

The Thesampling sampling function functionis isa a set setof ofimpulses impulses evenly evenlyspaced spaced on ona agrid. grid.

27
20 October 2011 1999-2011 by Richard Alan Peters II 23

The Sampling of an Image

An Animage imageis issampled sampled by bymultiplying multiplyingit itby by the thesampling samplingfunction function
I (r , c) samp N (r , c)

20 October 2011

1999-2011 by Richard Alan Peters II

24

Image Sampling in the Spatial Domain

samp N {I}(r , c) = I (r , c)

(r jN ) (c kN )
j = k =

Thisresults resultsin inan an This imagewhose whosepixels pixels image havethe thevalues valuesof of have theoriginal originalimage imageat at the theimpulse impulselocations. locations. the

20 October 2011

1999-2011 by Richard Alan Peters II

25

Convolution Property of the Fourier Transform

Let functions f ( r, c ) and g ( r, c ) have Fourier Transforms F (u, v ) and G (u, v ). Then,

F{ f g } = F G.
Moreover,

F { f g } = F G.
represents convolution represents pointwise multiplication

The Fourier Fourier Transform Transform of of The a product product equals equals the the a convolution of of the the Fourier Fourier convolution Transforms. Similarly, Similarly, the the Transforms. Fourier Transform Transform of of a a Fourier convolution is is the the product product convolution of the the Fourier Fourier Transforms Transforms of

20 October 2011

1999-2011 by Richard Alan Peters II

26

The Fourier Transform of the Sampling Function

The TheFourier Fourier Transform Transformof ofthe the sampling samplingfunction functionis is another anothersampling sampling function functionbut butwith with impulses /N impulses11 /Napart. apart.

Cf. slide 23
20 October 2011 1999-2011 by Richard Alan Peters II 27

Convolution by an Impulse
An impulse is a digital image, that has a single pixel with value 1; all others have value zero. An impulse at location (, ) is represented by:
( r , c ) =

1, if r = and c = 0, otherwise

If an image is convolved with an impulse at location (, ), the image is shifted in location down by pixels and to the right by pixels.

[I (r , c )] (r , c) = I (r , c ).
20 October 2011 1999-2011 by Richard Alan Peters II 28

Convolution by an Impulse

The convolution of any function with a delta function translates the function to the location of the impulse.
20 October 2011 1999-2011 by Richard Alan Peters II 29

The Fourier Transform of a Sampled Image

The The sampling samplingof of image imageII causes causesits its Fourier Fourier Transform Transform F{I} to tobe be repeated repeatedat at intervals intervalsof of 1/ N .. 1/ N

j = k =

F {I}(u, v ) u

j k v N N

If Ifthe the support supportof of F {I} has has a aradius radius N )) 1/(2 1/(2 N then then there therewill will be beoverlap overlap aliasing aliasing

20 October 2011

1999-2011 by Richard Alan Peters II

30

The FT of a Sampled Image with No Aliasing

For Foraliasing aliasingnot not to tooccur occurthe the frequency frequency support supportof ofII must musthave havea a radius N) radiusof of< <1/(2 1/(2 N)

j = k =

F {I}(u, v ) u

j k v N N

20 October 2011

1999-2011 by Richard Alan Peters II

31

Part of the Transform Computed by the FFT

as computed (aliasing)

origin shifted (aliasing)

as computed (no aliasing)

origin shifted (no aliasing)

Although the FT of the sampled image continues on indefinitely in theory, only one copy of the complete pattern is required for processing. The Fast Fourier Transform (FFT) algorithm computes the transform with the origin in the upper left. Usually the transform is displayed with the origin in the center.
20 October 2011 1999-2011 by Richard Alan Peters II 32

Filtering Before Downsampling to Prevent Aliasing

FT
To sample I(r,c) every N pixels multiply I(u,v) = F{I} by a window H(u,v) such that H (u, v) = 0 for u > 1 1 or v > 2N 2N Thus, to downsample I(r,c) by a factor of 2, use a spatial filter of width 4. Since this is midway between 3 and 5, use a 5x5 convolution mask. Ifthe theimage image If isaabit bitblurry blurry is thenaa33 will then xx 33will usuallysuffice suffice usually
33

That is equivalent to convolving I(r,c) with a spatial filter of width 2N.

20 October 2011 1999-2011 by Richard Alan Peters II

Filtering to Prevent Aliasing

original image
20 October 2011

power spectrum
1999-2011 by Richard Alan Peters II 34

Filtering to Prevent Aliasing

to decimate original by 2
20 October 2011

zero outside red square

35

1999-2011 by Richard Alan Peters II

Filtering to Prevent Aliasing

zoomed zoomed2 2
notice ringing in the result
20 October 2011 1999-2011 by Richard Alan Peters II

ideal filter
36

Filtering to Prevent Aliasing

To reduce ringing try multiplying the FT by a Gaussian w/ (v,u)=(R, C)

20 October 2011 1999-2011 by Richard Alan Peters II 37

Filtering to Prevent Aliasing

zoomed zoomed2 2
PS of FT Gaussian
20 October 2011

ringing is reduced but result is blurry

38

1999-2011 by Richard Alan Peters II

Filtering to Prevent Aliasing

try multiplying the FT

20 October 2011

by a Gaussian w/ (v,u)=(R, C)
39

1999-2011 by Richard Alan Peters II

Filtering to Prevent Aliasing

zoomed zoomed2 2
PS of FT Gaussian
20 October 2011

some aliasing but result is less blurry

40

1999-2011 by Richard Alan Peters II

Filtering to Prevent Aliasing

zoomed zoomed2 2
ideal filtered
20 October 2011 1999-2011 by Richard Alan Peters II

original
41

Filtering to Prevent Aliasing

zoomed 2 zoomed zoomed 2 zoomed 2 2

original
20 October 2011

Gaussian w/ (v,u)=(R, C)
1999-2011 by Richard Alan Peters II 42

Filtering to Prevent Aliasing

zoomed zoomed2 2 Gaussian w/ (v,u)=(R, C)

20 October 2011 1999-2011 by Richard Alan Peters II

original
43

Filtering to Prevent Aliasing

zoomed 2 zoomed zoomed 2 zoomed 2 2

original
20 October 2011

decimated by factor of 2
1999-2011 by Richard Alan Peters II 44