fourier transform in image...
TRANSCRIPT
-
Fourier Transform
6th Lecture on Image Processing
Martina Mudrová2004
in Image ProcessingJean Baptiste Joseph Fourier
1768-1830
-
Motivation
Why should we use the Fourier transformation?• basic mathematical tool to process images
image compression(format .JPG)
edge detection and image segmentation
quality image processing
image reconstruction
objects detection
etc.
2
Basic problems solved with the FT:
M. Mudrová, 2004
-
Mathematical Minimum
= transformation between time (spatial) x(t) and frequency X(f) domain
)()( fXtx ⇔ t…time domain (spatial)f…frequency domainBasic Definitions:
inverse continuous FT:direct continuous FT:
∫+∞
∞−
−= dtetxfX tfi π2 )()( ∫+∞
∞−
+= dfefXtx tfi π2 )()(
Requests to f(t) : continuous, including the 1st derivation
integrateable : ∞
-
Example of Continuous FT
How does look FT of periodical functions?
) 2sin()( 0tftx π= ( ))()( 21)( 00 ffffi
fX +−−= δδ
Euler’s formulas: )2sin()2cos( 002 0 tfitfe tfi πππ ±=±
∫+∞
∞−
−= dtetxfX tfi π2 )()(Def. FT:
⇔d(f)… Dirac impulse
ff0
Time (spatial) domain x(t) Spectrum |X(f)| (amplitude fr. charact.)
T=1/f0 2T 3T t
x(t)
-f 0 f f
|X(f)
|
0 0
⇔
4M. Mudrová, 2004
-
From Continuous to discrete FT
How is defined the discrete FT ?
Discretization in frequency: f Y fk Y k
Discretization in time: tY tn Y n +
Discrete Fourier TransformationDirect discrete FT: Inverse discrete FT:
)()( kXnx ⇔∑−=
−=
1
0
2 )(1)(
N
n
Nnki
enxN
kXπ
∑−
=
+=
1
0
2 )()(
N
k
Nkni
ekXnxπ
0 20 40 60-1
0
1
x(n)
n
|X(k
)|
0 0 0
31 3.14 0.5
15 1.57 0.25
k...indexωk ...circular frequencyf …normalized freq.
63
5M. Mudrová, 2004
-
Function Restriction in Time (Spatial) Domain
= apodizationFunction with definite length...
0 10 20 30 40 50 60
-1
0
1
x(n)
n
Periodický signál a jeho spektrum
0 0.5 1 1.5 2 2.5 30
10
20
30
40
ω od 0 do π
|X( ω
)|
0 10 20 30 40 50 600
0.5
1
1.5Obdelnikové okno a jeho spektrum
x(n)
n0 0.5 1 1.5 2 2.5 3
0
10
20
30
40
ω od 0 do π
|X( ω
)|
-1
0
1
x(n)
n
Omezený signál a jeho spektrum
0 0.5 1 1.5 2 2.5 30
10
20
30
40
ω od 0 do π
|X( ω)
|
6M. Mudrová, 2004
-
7
Let’s Go to Two Dimensions…
0 10 20 30 40 50 60-101
x(n)
n
n -> [n, m] k -> [k, l]
|X(k
)|
0 31 15 k 63
2D DFT
),(),( lkXmnx ⇔∑∑−=
−
=
+−=
1
0
1
0
)(2 ),(1),(
N
n
M
m
Mlm
Nkni
emnxMN
lkXπ
∑∑−
=
−
=
++=
1
0
1
0
)(2 ),(),(
N
k
M
l
Mml
Nnki
elkXmnxπ
Direct 2D DFT: Inverse 2D DFT:
1
6 4
1
6 40
1
n
x (n ,m )
m-1
01
-1
0
1
f k
| X ( f k , f l) |
f ln
m
1 6 4
1 64
x (n ,m )
M. Mudrová, 2004
-
Example of 2D DFT of Real Image
Symetry
0 1
1
0
1-1-1
1
n
m
1 366
1
270
x(n,m)
8M. Mudrová, 2004
-
Digital Filtering Principle
What happens in the case of spectrum processing?Spectrum |X(k)|Time space x(n)
1. Periodical function
2. Random noise
3. =1.+2.
4. After spectrum processing
9M. Mudrová, 2004
-
2D Filtering
10M. Mudrová, 2004
-
Discrete Convolution Principle
What is happening in the image space during the spectrum processing?Frequency space:
Multiplication
11
Time (image) space:Convolution
∑=
−==n
jjnhjxnhnxny
0)()()(*)()(
kkHkXkY pro )( . )()( ∀=
y(n)…requested functionx(n)...given (damaged) functionh(n) …digital filter
Where:
M. Mudrová, 2004
-
Discrete Convolution Example
What is a meaning of convolution?
12
∑=
−=
=n
jjnhjx
nhnxny
0)()(
)(*)()(
x(n) = 10; 20; 30; 40; 50h(n) = 0,5; 0,5
n=0 j=0 f(0).h(0) 10 . 0,5 5 y(0)= 5n=1 j=0 f(0).h(1-0) 10 . 0,5 5
j=1 f(1).h(1-1) 20 . 0,5 10 y(1)=5+10 = 15n=2 j=0 f(0).h(2-0) nedef.
j=1 f(1).h(2-1) 20 . 0,5 10j=2 f(2).h(2-2) 30 . 0,5 15 y(2)=10+15= 25
n=3 j=0 f(0).h(3-0) nedef.j=1 f(1).h(3-1) nedef.j=2 f(2).h(3-2) 30 . 0,5 15j=3 f(3).h(3-3) 40 . 0,5 20 y(3)=15+20= 35
n=4 j=0,1,2 nedef.j=3 f(4).h(4-3) 40 . 0,5 20j=4 f(4).h(4-4) 50 . 0,5 25 y(4)=25+20= 45
Example:(Sliding average)
y(n) = 5; 15; 25; 35; 45
M. Mudrová, 2004
-
13
2D Discrete Convolution
How does it seem the convolution in 2D?
∑∑= =
−−==n
j
m
jjmjnhjjxmnhmnxmny
0 02121
1 2
),(),(),(*),(),(
0 1 0
1 -4 1
0 1 0
filter matrixh(n,m)
image matrixx(n,m)
Element evaluatedin (n,m) position
M. Mudrová, 2004
-
Application I – Edge Detection
Goal: Edge detection
(=>image segmentation)
Orig
inal
Imag
eIm
age
afte
r im
age
dete
ctio
n
Edge = abrupt image function change
Principle:Application of edge detectors= highpass filters
14
Example of filtr spectrum |H(k,l)|Filter matrix example:h(n,m) (Prewitt filtr)
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−− 111000111
M. Mudrová, 2004
-
Application II – Noise Removing
Goal: Image Enhancement
Orig
inal
Imag
e
Principle:Application of low pass filter
Filter Spetrum |H(k,l)|
Pre
cess
edIm
age
15M. Mudrová, 2004
-
Application III – Sharpening
Goal: Corrupted image reconstruction
Cor
rupt
ed im
age
Rec
onst
ruct
ed im
age
Principle:
• Inverse filtering• Wiener filtering
Types of corruption:
• Defocussing due to object movement • Bad objective focussing• Atmosphere turbulentions• etc.
16M. Mudrová, 2004
-
Application IV – Object Detection
Goal: Finding the coordinates of a given object in the image
Principle:
17
• Evaluation of correlation function = convolution in the spatial domain
• Using an algorithm of FFT the computation gets faster-> multiplication in the
frequency domain
Imag
ex(
n,m
)
Filtr
h(n,
m)
Kor
elač
ní fu
nkce
h(n
,m)
M. Mudrová, 2004
-
Advanced Methods of Image Analysis
Short-time Fourier Transform- compromise between time (image)-frequency resolution
Wavelet transform-use time (image) window with various length- used in image analysis, denoising, compression
Radon transform-used for conversion from cylindric coordinate system-used mainly for biomedical image processing
…
18M. Mudrová, 2004
Fourier TransformMotivationMathematical MinimumExample of Continuous FTFrom Continuous to discrete FTFunction Restriction in Time (Spatial) DomainLet’s Go to Two Dimensions…Example of 2D DFT of Real ImageDigital Filtering Principle2D FilteringDiscrete Convolution PrincipleDiscrete Convolution Example2D Discrete ConvolutionApplication I – Edge DetectionApplication II – Noise RemovingApplication III – SharpeningApplication IV – Object DetectionAdvanced Methods of Image Analysis