chapter 4 image enhancement in the frequency...
TRANSCRIPT
Chapter 4
Image Enhancement in the
Frequency Domain
國立雲林科技大學資訊工程系
張傳育(Chuan-Yu Chang ) 博士
Office: EB 212
TEL: 05-5342601 ext. 4516
E-mail: [email protected]
Web: http://MIPL.yuntech.edu.tw
2
Background
Fourier series:
Any function that periodically repeats itself can be expressed as the sum of sines and/or cosines of different frequencies, each multiplied by a different coefficient.
Fourier transform
Even functions that are not periodic can be integral of sines and/or cosine multiplied by a weighting function.
A function expressed in either a Fourier series or transform, can be reconstructed completely via an inverse process, with no loss of information.
n
xT
nj
necxf
2
3
Background
Glass prism
The prism is a physical device that separates light into various
color components, each depending on its wavelength content.
The Fourier transform
The FT may be viewed as a “mathematical prism” that separates
a function into various components, based on frequency content.
The Fourier transform lets us characterize a function by its
frequency content.
4
Function s(x) (in red) is a sum of six sine functions of different amplitudes
and harmonically related frequencies.
Their summation is called a Fourier series.
The Fourier transform, S(f) (in blue), which depicts amplitude vs frequency,
reveals the 6 frequencies (at odd harmonics) and their amplitudes (1/odd
number).
Ref. Wikipedia (https://en.wikipedia.org/wiki/Fourier_series)
5
The 1D Fourier Transform
The Fourier transform, F(u), of a single variable, continuous
function, f(x), is defined as
where
We can obtain f(x) by means of the inverse Fourier transform
Extended to two variables, u and v.
dueuFxf uxj 2)()(
dxexfuF uxj 2)()(
dvduevuFyxf
dydxeyxfvuF
vyuxj
vyuxj
)(2
)(2
),(),(
),(),(
(4.2-1)
(4.2-3)
(4.2-4)
(4.2-2)
Fourier Transform Pair
(4.2-1) and (4.2-2)
A function can be recovered
from its transform.
1j
6
The 1D Fourier Transform (cont.)
Discrete Fourier Transform, DFT, of one variable
The concept of the frequency domain follows directly from Euler’s
Formula
Substituting (4.2-7) into (4.2-5), and using cos(-q )=cos( q ) obtain
1,...,2,1,0,)()(
1,...,2,1,0,)(1
)(
1
0
/2
1
0
/2
MxeuFxf
MuexfM
uF
M
x
muxj
M
x
muxj
qqq sincos je j
1
0
/2sin/2cos)(1
)(
M
x
MuxjMuxxfM
uF
(4.2-5)
(4.2-7)
(4.2-6)
(4.2-8)
在DFT時,乘數所放的位置並不重要,可放在轉換前或轉換後,或者兩個都放1/√M,但須滿足其乘積為1/M
每一分項稱為頻率成分(frequency component)
DFT:
IDFT:
Each term of the Fourier Transform is composed of
the sum of all values of the function f(x).
1,...,2,1,0, Mxxf
7
The 1D Fourier Transform (cont.)
As in the analysis of complex numbers, we find it
convenient sometimes to express F(u) in polar
coordinates
)()()()(
)(
)(tan)(
)()()(
where
)()(
222
1
22
)(
uIuRuFuP
uR
uIu
uIuRuF
euFuF uj
(magnitude, spectrum)
(phase angle, phase spectrum)
(power spectrum, spectral density)
Rear part of F(u)
imaginary part of F(u)
8
Example 4.1: Fourier spectra of two simple 1-D function
在x域曲線下的面積加倍時,頻譜上的高度也會加倍。(The height of the spectrum doubled as the area under the curve in the x-domain doubled.)
當函數長度加倍時,在相同區間上頻譜的零點加倍。(The number of zeros in the spectrum in the same interval doubled as the length of the function doubled)
A=1,K=8,M
=1024
u=0
9
The 1D Fourier Transform (cont.) In the discrete transform of Eq(4.2-5), the function f(x) for x=0, 1,
2,…,M-1, represents M samples from its continuous counterpart.
These samples are not necessarily always taken at integer values of x in the interval [0, M-1]. They are taken at equally spaced.
Let x0 denote the first point in the sequence. The first value of the sampled function is f(x0). The next sample has taken a fixed interval Dx units away to give f(x0+Dx). The k-th sample, f(k), is f(x0+kDx).
The sequence always starts at true zero frequency. Thus the sequence for the values of u is 0, Du, 2Du,…,[M-1]Du. The F(u)
Dx and Du are inversely related by the expression (in Fig. 4.2)
)()( 0 xxxfxf D
)()( uuFuF D
xMu
DD
1
(4.2-13)
(4.2-14)
(4.2-15)
10
Discrete-time Fourier transform
The discrete-time Fourier transform (DTFT) is
a form of Fourier analysis that is applicable to
the uniformly-spaced samples of a
continuous function.
The term discrete-time refers to the fact that
the transform operates on discrete data
(samples) whose interval often has units of
time.
11Depiction of a Fourier transform and its periodic summation (DTFT)
Ref. Wikipedia (https://en.wikipedia.org/wiki/Discrete-time_Fourier_transform)
12
The 2D DFT and Its Inverse
2D DFT pair The discrete Fourier transform of a function f(x,y) of size MxN is
given by
where u=0,1,2,…,M-1, v=0,1,2,…,N-1
The inverse Fourier transform is given by
1
0
1
0
)//(2),(1
),(M
x
N
y
NvyMuxjeyxfMN
vuF
),(),(),(),(
),(
),(tan),(
),(),(),(
222
1
22
vuIvuRvuFvuP
vuR
vuIvu
yxIyxRvuF
2D Fourier spectrum, phase angle, power spectrum
(4.2-16)
(4.2-17)
(4.2-18)
(4.2-19)
(4.2-20)
1
0
1
0
)//(2),(),(M
u
N
v
NvyMuxjevuFyxf
13
The 2D DFT and Its Inverse
It is common practice to multiply the input image
function by (-1)x+y prior to computing the Fourier
transform.
)2/,2/()1)(,( NvMuFyxf yx
Shifts the origin of F(u,v) to frequency coordinates (M/2,N/2),
which is the center of the MxN area occupied by the 2-D DFT.
This area of the frequency domain is called
Frequency rectangle
(4.2-21)
M,N須為偶數
14
The 2D DFT and Its Inverse (cont.)
The value of the transform at (u,v)= (0,0) is
F(0,0) sometimes is called the DC component of the
spectrum.
If f(x,y) is real, its Fourier transform is conjugate symmetric
Dx and Du are inversely related by
1
0
1
0
),(1
)0,0(M
x
N
y
yxfMN
F
),(),(
),(),( *
vuFvuF
vuFvuF
yNv
xMu
DD
DD
1,
1
(4.2-22)
(4.2-23)
(4.2-24)
(4.2-25)
(4.2-26)
The spectrum of the
Fourier transform is
symmetric.
15
Example 4.2:
Centered spectrum of a simple 2-D function
White rectangle of
size 20x40 pixels The image was multiplied by (-1)x+y
prior to computing the Fourier
transform
The image was processed prior
to displaying by using the log
transformation in Eq.(3.2-2)
16
Filtering in the Frequency domain
Some basic properties of the frequency domain
Frequency is directly related to rate of change.
The slowest varying frequency component (u=v=0)
correspond to the average gray level of an image.
As we move away from the origin of the transform, the low
frequencies correspond to the slowly varying components
of an image.
As we move further away from the origin of the transform,
the higher frequencies correspond to the faster and faster
gray level changes in the image.
17
18
The spectrum corresponding to the translated rectangle is identical to the
spectrum corresponding to the original image.
19
20
Example 4.3
An image and its Fourier spectrum
放大2500倍
21
In spatial domain
g(x,y)=h(x,y)*f(x,y)
In frequency domain
H(u,v) is called a filter.
The Fourier transform of the output image is
G(u,v)=H(u,v)F(u,v)
The filtered image is obtained simply by taking the
inverse Fourier transform of G(u,v)
Filtered Image=F -1[G(u,v)]
Filtering in the Frequency domain (cont.)
The multiplication of H and F
involves two-dimensional
functions and is defined on an
element-by-element basis.
22
Filtering in the Frequency domain (cont.)
Basics steps for filtering in the frequency domain
1. Multiply the input image by (-1)x+y to center the transform
2. Compute F(u,v), the DFT of the image from (1)
3. Multiply F(u,v) by a filter function H(u,v)
4. Compute the inverse DFT of the result in (3)
5. Obtain the real part of the result in (4)
6. Multiply the result in (5) by (-1)x+y
23
Some basis filters and their properties
Notch filter
It is a constant function with a hole at the origin.
Set F(0,0) to zero and leave all other frequency components
of the Fourier transform untouched.
otherwise
NMvuifvuH
1
2/,2/,0,
24
Example Result of filtering the image in Fig. 4.4(a) with a notch filter that
set to 0 the F(0,0) term in the Fourier transform.
In reality the average of the displayed image cannot be zero because the image has to have negative values for its average gray level to be zero and displays cannot handle negative quantities.
The most negative value was set to zero, and other values scaled up from that.
25
Some basis filters and their properties
Low frequencies in Fourier transform are responsible for the general gray-level appearance of an image over smooth areas.
High frequencies are responsible for detail, such as edges and noise.
A filter that attenuates high frequencies while passing low frequencies is called a low pass filter.
A lowpass-filtered image has less sharp detail than original image.
A filter that has the opposite characteristic is called a high pass filter.
A highpass-filtered image has less gray level variations in smooth areas and emphasized transitional gray-level detail.
Such an image will appear sharper.
26
Lowpass
Highpass
The effects of lowpass and highpass filtering
blurred image
Sharp image with
little smooth gray
level detail because
the F(0,0) has been
set to zero.
original image
27
Example
Result of highpass filtering the image in Fig. 4.4(a)
The highpass filter is modified by adding a
constant of one-half the filter height to the
filter function.
28
Correspondence between Filtering in the
Spatial and Frequency Domains
Convolution theorem
1
0
1
0
),(),(1
),(*),(
M
m
N
n
nymxhnmfMN
yxhyxf
1. Flipping one function about the origin.
2. Shifting that function with respect to the other by changing
the values of (x,y)
3. Computing a sum of products over all values of m and n,
for each displacement (x,y).
(4.2-30)
29
Correspondence between Filtering in the Spatial
and Frequency Domains (cont.)
Fourier transform pair
Impulse function of strength A, located at coordinates (x0, y0), is
denoted by Ad(x-x0, y-y0) and is defined by
A unit impulse located at the origin, which is denoted as d(x,y)
00
1
0
1
0
00 ,,),( yxAsyyxxAyxsM
x
N
y
d
),(*),(),(),(
),(),(),(),(
vuHvuFyxhyxf
vuHvuFyxhyxf
0,0,),(1
0
1
0
syxyxsM
x
N
y
d (位於原點的單位脈衝)
(4.2-31)
(4.2-32)
(4.2-33)
(4.2-34)
Used to indicate that the expression on
the left can be obtained by taking the
IFT of the expression on the right.
30
31
Correspondence between Filtering in the Spatial
and Frequency Domains (cont.)
The Fourier transform of a unit impulse at the origin
Let f(x,y)=d(x,y), carry out the convolution defined in
Eq.(4.2-30), using (4.2-34)
MN
eyxMN
vuFM
x
N
y
NvyMuxj
1
),(1
),(1
0
1
0
)//(2
d
),(1
),(),(1
),(*),(1
0
1
0
yxhMN
nymxhnmMN
yxhyxfM
m
N
n
d
(4.2-35)
(4.2-36)
只有在(x,y) = (0,0)時為1,其餘為0
在原點上的impulse Fouier Transform為實數值,相位角為0。如果impulse不在原點,則轉換後會產生非零的相位角(magnitude相同,但會平移)。
只有在(m,n) = (0,0)時為1,其餘為0
32
Correspondence between Filtering in the Spatial
and Frequency Domains (cont.)
Based on (4.2-31), combining (4.2-35) and (4.2-36)
Given a filter in the frequency domain, we can obtain the
corresponding filter in the spatial domain by taking the
IFT of the former.
We can specify filters in the frequency domain, take their
inverse transform, and then use the resulting filter in the
spatial domain as a guide for constructing smaller spatial
filter masks.
),(),(
),(),(),(),(
),(),(),(),(
vuHyxh
vuHyxyxhyx
vuHvuFyxhyxf
dd (4.2-37)
33
Introduction to the Fourier Transform
and the Frequency Domain (cont.) 高斯濾波器(Gaussian Filter)
The corresponding filter in the spatial domain is
These two equations represent an important result for two reasons: They constitute a Fourier transform pair, both components of
which are Gaussian and real. (上述二式構成Fourier transform pair,兩個成分均為高斯且為實數。)
These functions behave reciprocally with respect to one another. (這些函數彼此互為倒數。) 當H(u)有較大範圍的剖面時,h(x)有較窄的剖面。
22 2/)( uAeuH
22222)( xAexh
is the standard deviation
of the Gaussian curve
34
Introduction to the Fourier Transform
and the Frequency Domain (cont.)
We can construct a highpass filter as a difference of
Gaussians, as follows:
With A>=B and 1>2.
The corresponding filter in the spatial domain is
22
221
2 2/2/)( uu BeAeuH
222
2221
2 2
2
2
1 22)(xx
BeAexh
35
Introduction to the Fourier Transform
and the Frequency Domain (cont.)
We can implement
lowpass filtering in the
spatial domain by
using a mask with all
positive coefficients.
Once the values turn
negative, they never
turn positive again.
36
22
0
0
)2/()2/(),(
),(0
),(1),(
NvMuvuD
DvuDif
DvuDifvuH
Smoothing Frequency-Domain Filters Ideal Low-pass Filter
2 D ideal lowpass filter
Cuts off all high frequency components of the FT that are at a distance greater than a specified distance D0 from the origin of the transform.
(4.3-2)
(4.3-3)
從點(u,v)到傅立葉轉換中心點的距離
37
1
0
1
0
),(M
u
N
v
T vuPP
Smoothing Frequency-Domain Filters (cont.)
截止頻率(cutoff frequency)
H(u,v)=1和H(u,v)=0之間的過渡點。
整體功率(total image power)
Summing the components of the power spectrum at each
point (u,v).
百分比功率(% of the power)
u v
TPvuP /),(100
(4.3-4)
(4.3-5)
38
Example 4.4 Image power as a function of
distance from the origin of the DFT半徑為5, 15, 30,
80, and 230
功率比為92,
94.6, 96.4, 98,
and 99.5
39
Example 4.4 Image power as a function of
distance from the origin of the DFT (cont.)
存在振鈴現象(ringing)
40
Smoothing Frequency-Domain Filters (cont.)
對角線的profile
對角線的profile
41
Smoothing Frequency-Domain Filters (cont.) 巴特沃斯特低通濾波器(Butterworth Lowpass Filter)
BLPF transfer function does not have a sharp discontinuity (BLPF沒有銳利不連續的截止頻率)
Defining a cutoff frequency locus at points for which H(u,v) is down to a certain fraction of its maximum value. (將截止頻率定義在H(u,v)降到最大值的某個比例時。)
nDvuD
vuH2
0/),(1
1),(
42
Smoothing Frequency-Domain Filters (cont.)
2階巴特沃斯特低通濾波器,截止頻率半徑分別為5, 15, 30, 80,及230的結果。
43
Smoothing Frequency-Domain Filters (cont.)
The BLPF of order 1 has neither ringing nor negative values.
The filter of order 2 does show mild ringing and small negative values, but they certainly are less obvious than in the ILPF.
Ringing in the BLPF becomes significant for higher-order filter.
order 1 order 2 order 5 order 20
44
Smoothing Frequency-Domain Filters (cont.)
高斯低通濾波器(Gaussian Lowpass Filters)
Let = D0,
where D0 is the cutoff frequency.
22 2/,),( vuDevuH
20
2 2/,),(
DvuDevuH
45
Smoothing Frequency-Domain Filters (cont.)
Example 4.6: Gaussian
lowpass filtering
A smooth transition in
blurring as a function of
increasing cutoff
frequency.
No ringing in the GLPF.
46
Smoothing Frequency-Domain Filters (cont.)
A sample of text of poor resolution
Using a Gaussian lowpass filter with D0=80 to repair the
text.
47
Smoothing Frequency-Domain Filters (cont.) Cosmetric processing
Appling the lowpass filter to produce a smoother, softer-looking result from a sharp original.
For human faces, the typical objective is to reduce the sharpness of fine skin lines and small blemishes.
48
Smoothing Frequency-Domain Filters (cont.)
(a) High resolution radiometer image showing part of the Gulf of Mexico (dark) and Folorida (light). Existing many horizontal sensor scan lines.
(b) After Gaussian lowpass filter with D0=30.
(c) After Gaussian lowpass filter with D0=10. The objective is to blur out as much detail as possible while
leaving large features recognizable.
49
Sharpening Frequency Domain Filters
Edges and other abrupt changes in gray levels are
associated with high-frequency components.
Image sharpening can be achieved in the frequency
domain by a highpass filtering process.
Attenuating the low frequency components without
disturbing high-frequency information in the Fourier
transform.
The transform function of the highpass filters can be
obtained using the relation
vuHvuH lphp ,1, (4.4-1)
50
Sharpening Frequency Domain Filters
理想高通濾波器(Ideal Highpass Filters)
巴特沃斯高通濾波器(Butterworth Highpass Filters)
高斯高通濾波器 (Gaussian Highpass Filters)
0
0
),(1
),(0),(
DvuDif
DvuDifvuH
nvuDD
vuH2
0 ),(/1
1),(
20
2 2/,1),(
DvuDevuH
(4.4-2)
(4.4-3)
(4.4-4)
51
Sharpening Frequency Domain Filters (cont.)
It set to zero all frequencies inside
a circle of radius D0 while passing
without attenuation, all frequencies
outside the circle.
The IHPF is not physically
realizable with electronic
components, but it can be
implemented in a computer.
52
Sharpening Frequency Domain Filters (cont.)
Spatial representations of typical (a) ideal (b)
Butterworth, and (c) Gaussian frequency domain
highpass filters
53
Sharpening Frequency Domain Filters (cont.)
Result of ideal highpass filtering (a) with D0=15,
30, and 80
IHPFs have ringing properties.
54
Sharpening Frequency Domain Filters (cont.)
Result of BHPF order 2 highpass filtering (a) with
D0=15, 30, and 80
55
Sharpening Frequency Domain Filters (cont.)
Result of GHPF order 2 highpass filtering (a) with
D0=15, 30, and 80
56
Sharpening Frequency Domain Filters (cont.)
The Laplacian in the Frequency Domain
It can be shown that
Extended to two dimension
According to Eq.(3.7-1), the expression inside the brackets
on the left side of Eq.(4.4-6) is recognized as the Laplacian
of f(x,y). Thus, we have
uFju
dx
xfd n
n
n
vuFvu
vuFjvvuFjuy
yxf
x
yxf
,
,,,,
22
22
2
2
2
2
vuFvuyxf ,, 222
(4.4-5)
(4.4-6)
(4.4-7)
Laplacian
57
Sharpening Frequency Domain Filters (cont.)
Eq(4.4-7) presents that the Laplacian can be implemented in the frequency domain by using the filter
Assume that the origin of F(u,v) has been centered by performing the operation f(x,y)(-1)x+y prior to taking the transform of the image.
If f are of size MxN, this operation shifts the center transform so that (u,v)=(0,0) is at point (M/2, N/2) in the frequency rectangle.
The center of the filter function needs to be shifted:
The Laplacian filtered image in the spatial domain is obtained by computing the inverse Fourier transform of H(u,v)F(u,v)
22),( vuvuH
22 )2/()2/(),( NvMuvuH
),(2/2/,2212 vuFNvMuyxf
(4.4-8)
(4.4-9)
(4.4-10)
58
Sharpening Frequency Domain Filters (cont.)
Computing the Laplacian in the spatial domain using
Eq(3.7-1) and computing the Fourier transform of
the result is equivalent to multiplying F(u,v) by H(u,v).
The spatial domain Laplacian filter function obtained
by taking the inverse Fourier transform of Eq(4.4-9)
has some properties:
The function is centered at (M/2, N/2), its value at the top of
the dome is zero.
All other values are negative.
vuFNvMuyxf ,2/2/,222 (4.4-11)
59
Sharpening Frequency Domain Filters (cont.)
60
Sharpening Frequency Domain Filters (cont.)
We form an enhanced image g(x,y) by subtracting the
Laplacian from the original image
In Frequency domain, the enhanced image is obtained by
perform the operation with only one filter,
the enhanced image is obtained with a signal inverse
transform operation:
),(),(),( 2 yxfyxfyxg
),(2/2/1),(221 vuFNvMuyxg
(4.4-12)
(4.4-13)
222/2/1, NvMuvuH
61
Sharpening Frequency Domain Filters (cont.)
62
Unsharp masking, high-boost filtering
鈍化遮罩 (unsharp masking) 藉由減去影像自己的模糊化版本,所得到的銳化影像
藉由減去影像自己的低通濾波版本,以獲得高通濾波影像
High-boot filtering multiplying f(x,y) by a constant A>=1.
(4.4-15)可改寫成
將(4.4-14)代入,可得
),(),(),( yxfyxfyxf lphp
),(),(),( yxfyxAfyxf lphb
),(),(),()1(),( yxfyxfyxfAyxf lphb
),(),()1(),( yxfyxfAyxf hphb
(4.4-14)
(4.4-15)
(4.4-16)
(4.4-17)
63
Unsharp masking, high-boost filtering (cont.)
從上式的推導,鈍化遮罩可使用下列的複合濾波器直接在頻率域上實現
所以,high-boot filtering可以下列複合濾波器在頻率域上實現
),(1),(
),(),(),(),(
),(),(),(
),(),(),(
vuHvuF
vuFvuHvuFvuF
vuFvuHvuF
vuFvuFvuF
lp
lphp
lplp
lphp
),(1),( vuHvuH lphp (4.4-18)
),()1(),( vuHAvuH hphb (4.4-19)
64
Sharpening Frequency Domain Filters (cont.)
65
Sharpening Frequency Domain Filters (cont.)
66
Homomorphic filter Improving the appearance of an image by simultaneous gray-level
range compression and contrast enhancement.
An image f(x,y) can be expressed as the product of illumination and reflectance components:
Eq (4.5-1) cannot be used directly to operate separately on the frequency components of illumination and reflectance because the Fourier transform of the product of two functions is not separable
However, we define
Then
or
),(),(),( yxryxiyxf
yxryxiyxf ,,),(
yxryxi
yxfyxz
,ln,ln
,ln,
yxryxi
yxfyxz
,ln,ln
,ln),(
(4.5-1)
(4.5-2)
(4.5-3)
vuFvuFvuZ ri ,,, (4.5-4)
67
Homomorphic filter If we process Z(u,v) by means of a filter function H(u,v) then
from Eq.(4.2-27)
In spatial domain,
By letting
and
Eq. (4.5-6) can be expressed in the form
vuFvuHvuFvuH
vuZvuHvuS
ri ,,,,
,,,
vuFvuHvuFvuH
vuSyxs
ri ,,,,
,,
11
1
vuFvuHyxi i ,,, 1'
vuFvuHyxr r ,,, 1'
yxryxiyxs ,,, ''
(4.5-5)
(4.5-6)
(4.5-7)
(4.5-8)
(4.5-9)
68
Homomorphic filter (cont.) Since z(x,y) was formed by taking the logarithm of the original
image f(x,y), the inverse operation yields the desired enhanced
image, denoted by g(x,y)
where
The key to the approach is the separation of the illumination and
reflectance components. The homomorphic filter function H(u,v)
can then operate on these components separately.
),(),(
,
00
),('),('
,
yxryxi
ee
eyxg
yxryxi
yxs
),('0
),('0
),(
),(
yxr
yxi
eyxr
eyxi
(4.5-10)
(4.5-11)
(4.5-12)
69
Homomorphic filter (cont.)
Homomorphic filtering approach for image enhancement
The illumination component of an image generally is characterized by slow spatial variations, while the reflectance component tends to vary abruptly . Associating the low frequencies of the Fourier transform of
the logarithm of an image with illumination and the high frequencies with reflectance.
70
L
DvuDcLH evuH
20
2 /,1,
rH>1
rL<1
抑制低頻(照明),並放大高頻(反射) ,增加影像的對比度
Homomorphic filter (cont.) The HF requires specification of a filter function H(u,v) that
affects the low-and high frequency component of the Fourier transform in different ways.
The filter tends to decrease the contribution made by the low frequencies (illumination) and amplify the contribution made by high frequencies (reflectance). The net result is simultaneous dynamic range compression and
contrast enhancement.
71
Example: 4.10 In the original image
The details inside the shelter are obscured by the glare from the outside walls.
Fig. (b) shows the result of processing by homomorphic filtering, with L=0.5 and H=2.0.
A reduction of dynamic range in the brightness, together with an increase in contrast, brought out the details of objects inside the shelter.
72
Implementation Some additional properties of the 2D Fourier Transform
Translation properties:
when u0=M/2, v0=N/2
Eq(4.6-1) becomes
00//2
,, 00 vvuuFeyxfNyvMxuj
NvyMuxjevuFyyxxf
//200
00,,
yxyxjNyvMxujee
1
//2 00
2/,2/1, NvMuFyxfyx
vuvuFNyMxf
1,2/,2/
(4.6-1)
(4.6-2)
(4.6-3)
(4.6-4)
Same as Eq(4.2-21), which we used for centering the transform.
73
Implementation (cont.)
Distributivity
The Fourier transform is distributive over addition, but not
over multiplication
Scaling
yxfyxfyxfyxf ,,,, 2121
yxfyxfyxfyxf ,,,, 2121
bvauFab
byaxf
vuaFyxaf
/,/1
,
,,
(4.6-5)
(4.6-6)
(4.6-7)
(4.6-8)
74
Implementation (cont.)
Rotation
If we introduce the polar coordinates
x=r cos q, y=r sin q, u=w cos j, v= w sin j
then f(x,y) and F(u,v) become f(, q) and F(w, j)
Direct substitution into the definition of the Fourier transform
yields
The expression indicates that rotating f(x,y) by an angle q0
rotates F(u,v) by the same angle.
00 ,, qjwqq Frf (4.6-9)
75
Implementation (cont.)
Periodicity symmetry
Conjugate symmetry
NvMuFNvuFvMuFvuF ,,,,
NyMxfNyxfyMxfyxf ,,,,
vuFvuF
vuFvuF
,,
,, *
The complex conjugate of a complex number is the number with equal real part and imaginary part equal in
magnitude but opposite in sign. For example, the complex conjugate of 3 + 4i is 3 − 4i.
76
週期性表示F(u)有一長度為M的週期共軛對稱性表示頻譜的中心位於原點
Implementation (cont.)
77
Implementation (cont.) Separability
The discrete Fourier transform in Eq(4.2-16) can be
expressed in the separable form
where
1
0
/2
1
0
/21
0
/2
,1
,11
,
M
x
Muxj
N
y
NvyjM
x
Muxj
evxFM
eyxfN
eM
vuF
1
0
/2,1
,N
y
NvyjeyxfN
vxF
(4.6-14)
(4.6-15)
1
0
1
0
)//(2),(1
),(M
x
N
y
NvyMuxjeyxfMN
vuF
78
Implementation (cont.) Computing the inverse Fourier Transform Using a
forward transform algorithm 2D Fourier transforms can be computed via the application
of 1-D transforms.
The 1-D Fourier transform pair was defined as
Taking the complex conjugate of Eq(4.6-17) and dividing both sides by M yields
1,...,2,1,0,)(1
)(1
0
/2
MuexfM
uFM
x
Muxj
1,...,2,1,0,)()(1
0
/2
MxeuFxfM
x
Muxj
1,...,2,1,0,)(*1
)(*1 1
0
/2
MxeuFM
xfM
M
x
Muxj
(4.6-16)
(4.6-17)
(4.6-18)
Both (4.6-16) and (4.6-18) have same form
79
The complex conjugate of a
complex number is the number
with equal real part and imaginary
part equal in magnitude but
opposite in sign.
Ex: the complex conjugate of 3 + 4j is
3 − 4j.
Implementation (cont.)
80
Implementation (cont.)
Inputing F*(u) into an algorithm designed to
compute the forward transform gives the quantity
f*(x)/M.
Taking the complex conjugate and multiplying by M
yields the desired inverse f(x).
A similar analysis for two variables yields:
1
0
)//(21
0
),(*1
),(*1 M
x
NvyMuxjN
y
evuFMN
yxfMN
(4.6-19)
81
Implementation (cont.)
More on periodicity: the need for padding
Based on the convolution theorem, multiplication in the
frequency domain is equivalent to convolution in the spatial
domain, and vice versa.
Periodicity is part of the process, and it cannot be ignored.
1
0
)()(1
)(*)(M
m
mxhmfM
xhxf (4.6-20)
82
Each function consists of 400 points.
To mirror the function h(m)
about the origin.
To slide h(-m) past f(m).
By adding a constant x
At each displacement,
the entire summation in
Eq(4.6-20) is carried out.
Periods of the f(m) and h(m) extending
Infinitely in both direction.
Part of the first extended
period to the right of
h(x-m) lies inside the part
of f(m)
Failure to handle the periodicity issue
properly will give incorrect result
83
Implementation (cont.)
Assume that f and g consist of A and B points
We append zeros to both functions so that they have identical periods, denoted by P.
This procedure yields extended or padded functions
Unless we choose P>=A+B-1, the individual periods of the convolution will overlap.
If P=A+B-1, the periods will be adjacent.
If P>A+B-1, the periods will be separated.
PxB
Bxxgxg
PxA
Axxfxf
e
e
0
10
0
10(4.6-21)
(4.6-22)
84
Implementation (cont.)
85
Implementation (cont.)
If we wanted to compute the convolution in the
frequency domain, we would
Obtain the Fourier transform of the two extended sequences.
Multiply the two transforms
Compute the inverse Fourier transform.
86
Implementation (cont.)
Extensions to 2D function
Images f(x,y) and h(x,y) of sizes AxB and CxD
Wraparound error in 2-D convolution is avoided by
choosing
The periodic sequences are formed by extending f(x,y) and
h(x,y)
1
1
DBQ
CAP
QyDPxC
DyCxyxhyxh
QyBPxA
ByAxyxfyxf
e
e
and 0
10 and 10,,
and 0
10 and 10,,
Padding是必要的!因為在頻率域中,濾波的動作是將影像和濾波函數進行點乘的動作。
87
Implementation (cont.)
88
Implementation (cont.)
對圖4.12(c)以padded過的ideal low
pass filter,點乘後的結果。
對圖4.39 進行inverse Fourier
Transfer後的結果。
我們只須取出左上角的部分。
89
Implementation (cont.)
The Convolution and Correlation Theorems
兩函數f(x,y), h(x,y)的離散convolution可表示成
Convolution和Fourier transform間的關係
兩函數f(x,y), h(x,y)的離散correlation可表示成
1
0
1
0
),(),(1
),(*),(M
m
N
n
nymxhnmfMN
yxhyxf
),(),(),(*),( vuHvuFyxhyxf
),(*),(),(),( vuHvuFyxhyxf
1
0
1
0
* ),(),(1
),(),(M
m
N
n
nymxhnmfMN
yxhyxf
(4.6-27)
(4.6-28)
(4.6-29)
(4.6-30)
對於空間域的影像而言,由於沒有complex部分,所以f*=f,
90
Implementation (cont.)
空間與頻率上的correlation
Correlation可用在matching,假設f(x,y)是包含某些物件的影像, h(x,y)為template,如果有匹配的話,兩函數的correlation會有最大的值。
),(),(*),(),( vuHvuFyxhyxf
),(),(),(),(* vuHvuFyxhyxf (4.6-32)
(4.6-31)
91
Implementation (cont.)
Correlation
Cross correlation
The image being correlated are different.
Autocorrelation
Both images are identical.
The autocorrelation theorem
the Fourier transform of the spatial autocorrelation is
the power spectrum.
Similarly, according to Eq.(4.6-32), we have
2),(),(),( vuFyxhyxf
),(),(),(2
vuFvuFyxf (4.6-34)
(4.6-33)
92
Implementation (cont.)
256
256
42
38
256+42-1=297
256+38-1=293Padded images
最大的correlation
value出現在”T”上面。
93
Chapter 4Image Enhancement in the
Frequency Domain
94
Chapter 4Image Enhancement in the
Frequency Domain
95
Chapter 4Image Enhancement in the
Frequency Domain
96
Chapter 4Image Enhancement in the
Frequency Domain
97
Chapter 4Image Enhancement in the
Frequency Domain
98