p. 1 some important noise models p. 2 lt 10 lecture 10 · noise models and noise estimation ......
TRANSCRIPT
Digital Image Processing L t 10
p. 1
Lecture 10 Noise reduction Noise reduction
Noise models and noise estimation Using linear filters, ex) average filters Using non linear filters ex) the median filter Using non-linear filters, ex) the median filter Using adaptive filters (spatial variant) Periodic Noise Reduction by Frequency domain filtering
Image restoration Image restoration The degradation model Inverse filtering Least square filtering Least square filtering Wiener filtering Estimating the degradation function An often cited standard method: Richardson–Lucy deconvolution An often cited standard method: Richardson Lucy deconvolution
(not in Gonzalez & Woods) (Extra, optional) Gonzales & Woods:
Chapter 5, pp. 311-362 Paperback 2013: Add 22 to page numbers Chapter 5, pp. 311 362 Paperback 2013: Add 22 to page numbers Chapter 4, pp. 294-298 Paperback 2013: Add 22 to page numbers
Maria Magnusson, CVL, Institutionen för Systemteknik, Tekniska Högskolan, Linköpings Universitet
Some important noise modelsd th i i ti
p. 2
and their use in practice Gaussian noise Gaussian noise
Easy to use mathematically. Therefore often used in practice even if the model is not perfect.
Electronic circuit noise and sensor noise due to poor *
pillumination* or high temperature. *= rather poisson noise
Rayleigh noise Occurs in range imaging. Can model skewed histograms.
Gamma and Exponential noise Occurs in laser imaging. Can be used for approximating
skewed histograms.U if i Uniform noise Not so practical, but can be useful in random number
generation in simulations. Salt-and-pepper (impulse) noise Salt-and-pepper (impulse) noise
Quick transients due to as faulty switching during imaging Periodic noise
Electrical or electromechanical interference during image Electrical or electromechanical interference during image acquisition
Newspaper printing
S i t t i d l
p. 3
Some important noise models
Fig. 5.2Eq. (5.2-1)-(5.2-14)
A simple imageith diff t i
Fig. 5.4
p. 4
with different noise
A simple imageith diff t i
Fig. 5.4
p. 5
with different noise P i i
p. 6
Poisson noise probability mass function: probability mass function:
,...1,0för,!
Pr kekakN a
k
Mean and variance:!k
aa 2and Is used to statistically characterize the distribution of
photons per unit of area
aa NN and
photons per unit of area. Important for normal camera images, and medical image
systems using X-rays and gamma-rays.systems using X rays and gamma rays. Also called Poisson noise, Photon noise or Quantum noise.
P i i
p. 7
Poisson noise Previous slide gave that Poisson noise has equal mean
and variance:It b i t d ith G i i ith
aNN 2 It can be approximated with Gaussian noise with
aNN 2
P i i
p. 8
Poisson noise The standard
deviation
i ith th
aN
increase with the signal value
B t the signal to
aN
But the signal to noise ratio
improve! aaa
M di filt t h i
p. 9
Median filter, techniqueTh di filt The median filter moves over the in-image, similarly as during convolution.du g co o ut o
The values in the in-image, under the filter, are noted. E l 1 1 9 2 3 3 3 3 3Example: 1 1 9 2 3 3 3 3 3
The values are sorted to: 1 1 2 3 3 3 3 3 91 1 2 3 3 3 3 3 9
The median value becomes 3 in the example. It is put in Image with movingp pthe out-image.
g gmedian filter
Sorting is gcomputationally
heavy!
Median and average filter i
p. 10
comparison
Original WithOriginal image
With salt- andpepper
inoise
AfterAfter medianfilter3x3
averagefilter 3x3 3x33x3
P ti f th di filtp. 11
Properties of the median filter Edges are
preserved.N i i d Noise is suppressed (especially salt-and-pepper noise)and pepper noise).
Thin lines are destroyed.destroyed.
Smooth surfaces arise.
The median filter b li d l ti
p. 12
can be applied several times
Image AfterImage with salt-and-pep-
i
Aftermedianfilter 3x3,1 tiper noise 1 time
AfterAftermedianfilter 3x3,3 times
medianfilter 3x3,2 times
Fig. 5.10
3 times2 times
Median and average filter i
p. 13
comparison Lab 7
Image Salt andImage with additive niform
Salt- and pepper noisedd duniform
noiseadded
After AfterAfteraveragefilter 5x5
Aftermedianfilter 5x5
Fig. 5.12
Adaptive filtering p. 14
(wiener2 according to MATLAB)C t d i f i N2 i th Compute mean and variance of size N2 in the neighborhood :
2f f 2
,
2
2
,
N
yxfyx
2
,,
N
yxfyx ≈(5.2-15) ≈(5.2-16)
The adaptive filtering expression:
NN
p g p
yxfyxg n ,, 2
22(5.3-12)
where is the variance of the noise.
yfyg ,, 2
2n
Adaptive and average filter i
p. 15
comparison Lab 7
ImageImage with additive Ga ssian
Afteraveragefilter 7x7Gaussian
noisefilter 7x7
AfterAfteradaptivefilter 7x7
Fig. 5.13
Periodic noise reduction by f d i filt i
p. 16
frequency domain filtering Fig. 5.16
Image Fou-Image with sinus-oidal
Fou-rier trans-formoidal
noiseform
Re-Butter-sult afterfilte-
worthband-reject te
ringeject
filter
B d j t filt ith ti
p. 17
Bandreject filters with equations
Fi 5 15Fig. 5.15
Tab. 4.6
OBS!
Notch (reject) filtersill t t d
p. 18
illustrated
Fig. 5.18
Notch (reject) filtersith ti
p. 19
with equations OBS!
( ) T b 4 4
,,,Q
kkNR vuHvuHvuH
(4.10-2):be shouldh Butterwort Lowpass
Tab. 4.4
,and,where
,,,1
kk
kkkNR
vuHvuH
vuHvuHvuH
,),(1
1,2
vuDvuH
n
lyrespectiveandat centered filters highpass are
kkkk vuvu radialtheiswhere
),(10
vuD
D
lyrespective, and , kkkk vuvu
Tab. 4.5
center thefrom distance
radial theis , where vuD
0 +101
Ex) Notch filt
p. 20
filtersFourier
S t llit
Fouriertransform
NotchSatelliteimage of Florida
Notch pass filter
Result after notchnotch reject filter Result
afterafter notch pass filter
Fig. 5.19
filter
Ex) Notchfilt
p. 21
filters
Sampled newsnews-paperimage
Fouriertransform
Result after
Notch reject after
notch reject filt
reject filter applied on the filter
Fig. 4.64
on the Fourier transform
Image restoration and I h t
p. 22
Image enhancement Image restoration
To recover a degraded image An objective process An objective process Uses a priori knowledge of the degradation phenomenon Example: Removal of image blurp g
Image enhancement To make an image easier to study A heuristic process Take advantage of the psychophysical aspects of the
human visual systemhuman visual system Example: Contrast stretching
Th i t ti bl
p. 23
The image restoration problem An image has often been exposed to a known
degradation.
vuNvuFvuHvuGyxnyxfyxhyxg ,,,, (5.1-1)
(5 1-2) vuNvuFvuHvuG ,,,,
Noise:
(5.1-2)
g(x,y)G( )+
n(x,y), N(u,v)
Degradation:h(x y) H(u v)
f(x,y)F( ) G(u,v)h(x,y), H(u,v)F(u,v)
We wish to find out as much as possible about f(x,y).
Th i t ti d l
p. 24
The image restoration model We know h(x,y), H(u,v) and statistical properties
for n(x,y), N(u,v).Noise:
g(x,y)+
Noise:n(x,y), N(u,v)
Degradation:f(x,y) g(x,y)G(u,v)
+h(x,y), H(u,v)f(x,y)F(u,v)
Restaurationfilter
vuF
yxf,ˆ,ˆ
Fig. 5.1
We look for that are as close as vuFyxf ,ˆ,,ˆ
vuF , g
possible to f(x,y), F(u,v). yf ,,,
I t ti l
p. 25
Image restoration, example( )h( )
Lab 7
n(x,y)h(x,y)
+*
( )f(x,y)
g(x,y)
Restorationfilter
yxf ,ˆ
T ti l bl
p. 26
Two partial problems Find the model (i.e. h(x,y) and n(x,y)) that best
describes the degradationFi d th t ti filt th t b t t th Find the restoration filter that best restores the original image.
N i
g(x y)+
Noise:n(x,y), N(u,v)
Degradation:f(x y) g(x,y)G(u,v)
+Degradation:h(x,y), H(u,v)
f(x,y)F(u,v)
Restaurationfilter
vuF
yxfˆ
,ˆ
vuF ,Fig. 5.1
Th t ti th d
p. 27
Three restoration methods 1) Inverse filtering 1) Inverse filtering 2) Constrained least squares filtering 3) Wiener filtering 3) Wiener filtering
(minimum mean square error filtering)
ExamplesExamples of images that can be restoredof images that can be restored Images with out-of-focus blur Images with motion blur Images with motion blur Images distorted by with atmospheric blur Some microscopy images, especially the z-py g p y
direction for 3D confocal microscopy
T t i
p. 28
Test imageh d d d h The test image is an aerial image degraded with
atmospheric turbulence. (5.6-3)
Degradation model: Fig. 5.25
Test image Degraded image, k=0.0025 Degraded image, k=0.001
1) Inverse filtering anddifi d i filt i
p. 29
-modified inverse filtering
vuGvuF ,,ˆ (5 7-1) vuHvuF
,, (5.7 1)
vuGF ,ˆ
vuH
vuF,
,,
yxnyxfyxhyxg ,,,,Compare
vuNvuFvuHvuG ,,,,with:
R lt ith I filt i
p. 30
Result with Inverse filteringI fil i li d h d d d i k 0 0025 Inverse filtering applied to the degraded image, k=0.0025.
For high frequencies, the noise dominates over the signal. Therefore the inverse filter is preferably limited to a radius e e o e e e se e s p e e ab y ed o a ad usless than W.Here: W=40, W=70 and W=“the band limit”.
Fig. 5.27
Full inverseCut off outside 40 Cut off outside 70
2) Constrainedl t filt i
p. 31
least squares filtering
22
* ,,,ˆ vuGvuHvuF (5 9-4)
22 ,,,
vuPvuH (5.9 4)
Where |P(u,v)| is the absolute value of the Laplacian filter, i.e. 2224P Laplacian filter, i.e.
or P(u,v) is the Fourier
2224, vuvuP
010 or P(u,v) is the Fouriertransform of
141010
, yxp (5.9-5)
010
, yp
Proof for Constrainedl t filt i
p. 32
least squares filteringTh b diff F h i i i I( ) b l
vuFHGvuI ˆ 2
There can be many different F that minimizes I(u,v) below:
vuFHGvuI ,,
The negative Laplace filter enhances high frequencies:
2224 vuvuP
The F that minimize the equation below cannot
4, vuvuP
ˆˆ 22
The F that minimize the equation below cannot contain too much high frequencies:
vuFPFHGvuI ,,
Proof for Constrainedl t filt i
p. 33
least squares filtering cont.Mi i i hi
Equivalent
vuFPFHGvuI ˆˆ 22
Minimize this eq:q
equations. Willbe shown on the
white board if vuFPFHGvuI ,,
Rewrite it to:
white board if we have time.
vuGHGHFvuI 1ˆ 2
22*
vuGFvuI ,1,
This gives the constrained
l t filt i22 PH least squares filtering
equation directly!
Example) frequency response for p. 34
constrained least squares filtering
1
H 0,1uH
ku
vuHsinc
,
0,uH kvsinc
22
* 0,uH
,
22 0,0, uPuH
It ti d t i ti f
p. 35
Iterative determination of f f 1) Specify an initial value of
2) Compute2 2
F̂HG ≈(5.9-6) and (5.9-7)
3) Stop if is sufficiently small. Otherwise chose a new and go to 2.
Result with constrainedl t filt i
p. 36
least squares filtering
Fig. 5.31Fig. 5.25
Not optimal Iteratively determined Degraded test image
3) Wi filt i
p. 37
3) Wiener filtering
där,,,ˆ vuWvuGvuF
vuHvuW ,, 2
*
(5.8-2)
vuSvuSvuH FN ,,, 2
whereis the power spectrum of the un degraded image
vuSvuS NF ,and,is the power spectrum of the un-degraded image and noise, respectively.
Note that vuSvuS NF ,,is proportional to the SNR (signal-to-noise ratio).
NF ,,
Wh Wi filt ?
p. 38
Why Wiener filter?Th Wi filt i li Norbert Wiener The Wiener filter is a linear filter that is optimal in the sense that it minimizes the
Norbert Wiener, USA,1894-1964se se t at t es t e
mean square error between the undistorted image f(x,y) and the restored imageand the restored image yxf ,ˆ
i.e. it minimizes
22 ˆ
2 ffEe
Autocorrelation and t
p. 39
power spectrum
dbdabyaxfbafyxrf
,,,
The autocorrelation function in the Fourier domain:in the Fourier domain:
vuSvuFvuFvuFyxr Ff ,,,,, 2* y Ff ,,,,,
The autocorrelation function and theThe autocorrelation function and the power spectrum is a Fourier pair:
j 2
dydxeyxrvuS yvxujfF
2,,
Autocorrelation, power spectrum p. 40
and stochastic signals (ex) noise)Thi i lid f t h ti i l
byaxnbanEyxrn ,,,This is valid for stochastic signals:
yyn ,,,
2 xn xrnWhite noise:
2uNuSN
By using the autocorrelation function and the power
x x u
By using the autocorrelation function and the power spectrum, we can make calculations with noise!
(Example of a li ti i )
p. 41
more realistic noise)
Bandlimited white noise:
xn xrn 2uNuSN
f
x x u
1
1f
xn efxr 1
21
21
2
11
fuuNuSN
1f
Example) frequency response f t Wi filt
p. 42
for two Wiener filters
vuH , kusinc,
kvsinc
dB10,0, SNRuW
dB200 SNRW dB20,0, SNRuW
Result with approximate Wiener filt i i t ti l h K
p. 43
filtering, interactively chosen K H *
KvuH
vuHvuW
2
*
,,, (5.8-6)In Lab 7, we will
use the non-approximate
Fig. 5.28
KvuH ,Wiener filter
Inverse, cut off outside 70 Wiener Filter, (5.8-6)Degraded test image
What do we need to know to be bl t f Wi filt i ?
p. 44
able to perform Wiener filtering?Th t l ti f ti f th i ( ) The autocorrelation function of the noise, rn(x,y) Ex) rn(x,y)=n2 (x,y) for white noise with variance n2
The autocorrelation function of the image, rf(x,y)Ex) 22 Ex) 222, yx
ff yxr
The degradation function h(x,y) Ex) the Gauss-like function that models atmospheric
turbulenceturbulence
The autocorrelation functionf th i ( )
p. 45
of the noise, rn(x,y) The noise can be measured here. Suppose that
the noise is white so that rn(x,y)=n2(x,y). The variance 2 can be calculated here in the square variance n2 can be calculated here in the square of size N2:
22
, yxfyx
2,2
Nyx
n ≈(5.2-16)
,
, yxfyx
where
2,
Nyx ≈(5.2-15)
The autocorrelation function of th i ( )
p. 46
the image, rf(x,y)
byaxfbafyxrf ,,,a b
f f bafyxfyxfyxr ,,,, 2
Th t l ti f ti i l i l i
a b
f
The autocorrelation function is always maximal in the point (x,y)=(0,0). This is valid for all images.
The autocorrelation function is always The autocorrelation function is always symmetrical, i.e. rf(x,y)=rf(-x,-y)
The autocorrelation function f th i ( )
p. 47
of the image, rf(x,y) Lab 7
This model is good for natural images:
222 yxff yxr
2f , ff yxr
yf
x It is supposed that It is supposed that
images of the same type, for yp ,example faces or license plates, h th have the same
The autocorrelation function f i ( )
p. 48
of one image, rf(x,y) Lab 7
Circular convolution effects in
222, yxff yxr
Circular convolution effects in the outer area of the image because the computation was
f d i th F iff
performed in the Fourier domain.
3 ways to estimate the degradation p. 49
function H(u,v) 1) Estimation by image observation
If the image is blurred, we can look at a small rectangular section of the image containing sample rectangular section of the image containing sample structures, like part of the object and the background
Look at areas with high contrast to diminish the effect of inoise
Compute
vuG vuF
vuGvuHs
ss ,ˆ
,, (5.6-1)
where gs(x,y) is the observed sub-image and fs(x,y) is the estimated original image in the actual sub-area
3 ways to estimate the degradation p. 50
function H(u,v) Estimation by experimentation
Show an impulse image to the system and observe an image g(x y)image g(x,y).
Compute
vuGvuH ,, (5.6-2)
where G(u,v) is the Fourier transform of the observed image and A is the strength of the impulse
A
vuH , (5.6 2)
image and A is the strength of the impulse. Estimation by modeling
Ex1) Use the model for atmospheric turbulence Ex1) Use the model for atmospheric turbulence Ex2) This is valid for linear motion
vbuajbTH i vbuajevbuavbua
vuH
sin, (5.6-11)
System identification & image restoration p. 51
with constrained least squares filteringa) Image restoration b) System identification
n(x,y)a) Image restoration
? n(x,y)b) System identification
?g(x,y)
+h(x,y)f(x,y) g(x,y)
+h(x,y)f(x,y)
?
+h(x,y)
Mi i i
+h(x,y)
Mi i iMinimize:
22 ˆˆ FLFHG
Minimize:
22 ˆˆ HLFHG FLFHG HLFHG *
ˆ GHF *ˆ GFH
22 LHF
22 LFH
An often cited standard method:p. 52
Extra,
Richardson–Lucy deconvolutionTh Ri h d L l i h l k L
optio-nal
The Richardson–Lucy algorithm, also known as Lucy–Richardson deconvolution, is an iterative procedure for recovering a latent image f(x,y) that has been blurred by a known point spread function, h(x,y).
The observed image is denoted g(x,y). The statistics are performed under the assumption that are The statistics are performed under the assumption that are
Poisson distributed, which is appropriate for photon noise in the data:
yxh
hfyxgyxfyxf jj ,ˆ
,,ˆ,ˆ1
The start image f^0 can be a constant image.
yxhyxf j
jj ,,
g 0 g Matlab experiments on next page: original, blurred and
noisy, restorated with 5 iterations, and 15 iterations.