L. Yaroslavsky. Course 0510.7211 “Digital Image Processing: Applications” Lecture 11. Image restoration and enhancement: Linear filters
Signalling and imaging devices, signal distortions and the inverse problem.
Least squares restoration: AVERR AV AV a aim sys im par kid
krst
k
N
= −⎛⎝⎜
⎞⎠⎟=
−
∑. .
2
1
1
.
Linear filters: ( )A Arst inp= LF .
Filtering in transform domain. Scalar filters: ( )A T H T Arstd
inp= −1 ; inprr
rstr αηα = ;⇒
( )( ) ⎟⎠⎞⎜
⎝⎛=
∗ 2inprpar.imsis.im
inpr
idrpar.imsis.im
optr AVAV/AVAV αααη
Scalar filters for suppressing additive
noise: A A Ninp id= + ⇒2
.
2
.
2
.
rsysimidrparim
idrparim
optr
AVAV
AV
νγα
αη
+⎟⎠⎞⎜
⎝⎛
⎟⎠⎞⎜
⎝⎛
=
Empirical Wiener filters
( )( ) ⎥
⎥
⎦
⎤
⎢⎢
⎣
⎡ −≈ 2
22
0inprparim
rsysiminprparimopt
rAV
AVAV
α
ναη
.
..,max .
Signal power spectra estimation problem
Rejecting filters (transform shrinkage):ηα
αr
im par rinp
im par rinp
AV thr
AV thr=
⎛⎝⎜
⎞⎠⎟ ≥
⎛⎝⎜
⎞⎠⎟ <
⎧
⎨⎪
⎩⎪
1
0
2
2
,
,
.
.
. Soft and hard thresholding.
Image deblurring: A LA N Linp id= + , where isa linear operator α λ α νrinp
r rid
r= +
( )
( )ηλ
λ α
λ α ν λropt
r
im par r rid
im par r rid
im sys r r
r
r
AV
AV AV
SNRSNR
=+
=+
1 11
2 2
2 2 2
.
. .
Applicability and drawbacks of Wiener-type filtering. Application examples: filtering periodic noise; filtering stripe-noise. Wavelet (sub band decomposition) shrinkage. Multi component signal restoration Least squares approach { } ;K,...,k;NALA k
idkk
inpk 1=+=
( );;
.
,.
,,,
∑∑
=
•
= +== K
mm
l
idlparim
idl
idkparim
l
optlk
K
l
inplk
rstk
SNR
SNR
AV
AVS
AHA
1
21 1
1
α
ααη
( )( ) ⎪⎭
⎪⎬⎫
⎪⎩
⎪⎨⎧
= 2
22
lsysim
idlpariml
lAV
aAVSNR
ν
λ
.
.
Potential restoration quality: ∑=
=K
llSNR/
energysignalerrornrestoratio.rms
11
Super resolution from multiple images. Correlational averaging (see also Lect. 9). Local adaptive linear filters
Local criteria: ( ) ( ) ( )AVLOSS k l AV LOC m n k l LOSS a astat m n m nm n
, , / , ∃ ,, ,,
=⎧⎨⎩
⎫⎬⎭
∑
Spatial neighbourhood. Moving window linear local adaptive filtering in transform domain. Transform selection. Recursive spectral analysis and filtering in DFT, DCT and Haar domains and its computational complexity. 3-D Local adaptive filtering for restoration of multi component images. Problems for self-testing: 1. Formulate the least squares signal restoration problem. Describe scalar filtering in transform
domain and its advantages and limitations.. 2. Describe local adaptive filtering methods and explain their computer implementation. Why DCT is
advantageous to other transforms in this application? 3. Explain approaches to multi component signal restoration and super resolution. 4. Let a set of images of the same object is available. How this can help to improve image quality?
A canonical block-diagram of imaging systems
MSE-OPTIMAL (WIENER) FILTERING: FILTERING NARROW-BAND NOISE
Input image
Filtered image
Filtering periodic noise
Noisy image (scanning atomic force microscope image)
Filtered image
Filtering “banding” noise
20 40 60 80 100 120
6
8
10
12
14Column wise averaged power spectrum along rows
20 40 60 80 100 120
6
8
10
12
14"Filtered" power spectrum; thr=0.01
50 100 150 200 250 90
110
130
150
Average along rows versus row number
50 100 150 200 250
50
100
150
200
250Filtered averaged rows
Linear transformation
Point-wise nonlinear
transformation ( )xa ( )xb
Stochastic transformation
Wiener filtering for noise suppression
Noisy image (Wiener.m, RD=1000; SNR=2)
Noise, std=127.8706
Ideal Wiener filter
Restoration error, std=55.3361
Empirical Wiener filter, 1=γ
Restoration error, std=68.3752
Image processing: ”local vs global” Justification of local processing and adaptation:
It is well known that, when viewing image, human eye’s optical axis permanently hops chaotically over the field of view(A. Yarbus, Eye Movements and Vision, Plenum Press, New-York, 1967. Translated from the Russian edition (Moscow, 1965)) and that the human visual acuity is very non-uniform over the field of view. The field of view of a man is about 30°. Resolving power of man's vision is about 1′. However such a relatively high resolving power is concentrated only within a small fraction of the field of view that has size of about 2° (M.D. Levine, Vision in Man and Mashine, McGraw-Hill, 1985, pp. 110-130]). Therefore, area of the acute vision is about 1/15-th of the field of view.
Test image (left) and results of recording eye fixation when observing this image (right)
Image frame
Image global DCT-spectrum
Image blocks
DCT spectra of blocks
Fragmentation of the image into blocks demonstrates spatial inhomogeneity of the image and the fact that the structure of individual blocks is much simpler than that of the whole image. Adaptive filter design assumes empirical evaluation of signal statistical parameters such as spectra (for local adaptive linear filters). In global image statistics, parameter variations due to image non-homogeneity are hidden and are difficult if not impossible to detect. Therefore in global statistical analysis image local information will be neglected in favour of global one, which usually contradicts processing goals.
LOCAL ADAPTIVE FILTERING IN TRANSFORM DOMAIN
Local adaptive filtering in DCT domain: suppressing additive noise in electrocardiogram
5 0 0 1 0 0 0 1 5 0 0 2 0 0 01 0 0 0
1 5 0 0
2 0 0 0
2 5 0 0
In i t i a l s i g n a l
L o c a l D C T s p e c t r u m ;S z W = 2 5
5 0 0 1 0 0 0 1 5 0 0 2 0 0 0
51 01 52 0
2 5M o d i f i e d lo c a l s p e c t r u m ; S z W = 2 5 ; T h r = 0 . 0 0 1
5 0 0 1 0 0 0 1 5 0 0 2 0 0 0
51 01 52 0
2 5
5 0 0 1 0 0 0 1 5 0 0 2 0 0 0
1 5 0 0
2 0 0 0
2 5 0 0
R e s t o r e d s i g n a l
Transform
Scanning direction
Transform domain
Filter formation
Filter
Point-wise multiplication
by the filter mask
Modified DCT
spectrum
Inverse transform
Output pixel
Output image
Input image
Local adaptive filtering in DCT domain: image denoising
Initial (upper left), noisy (upper right), filtered (bottom left) images and filter
“transparance” map (bottom right)
Noisy image (frgm+40*(rand(256)-0.5));
Filtered image (out = lcdct_trsh(9,9,150,frgm_n));
Filter “transparance” map
Local adaptive filtering: 3D (space-colour) denoising colour images
Local adaptive filtering: blind image restoration (initial and restored images)
Filtering speckle noise in Ultra Sonic images: original (top) and filtered I(bottom) images
Test images
Noisy images
2-D DCT domain de-noising in window 5x5
3-D DCT domain de-noising in window 5x5x5
Local adaptive 2-D and 3-D DCT domain filtering for image de-noising
Frames of a movie acquired by a thermal camera: initial image (top) and filtered image, 3-D DCT domain 5x5x5 window, (bottom)
Local adaptive filtering in DCT domain: Image sharpening and local contrast enhancement
Initial and enhanced air-photograph (lcdct2(aero512,ones(9),0.75,2.5,1.5));
Filtering in transform domain: wavelet shrinkage (D.L. Donoho and I.M. Johnstone, Ideal spatial adaptation by wavelet shrinkage,
Biometrica, 81(3): 425-455, 1994)
Low pass filtering and
downsampling Interpolation
-+ + Soft/hard
thresholding
Interpolation
+ + +
+ + +
Input
Output
Interpolation
-
+
+
Interpolation
Low pass filtering and
downsampling
Low pass filtering and
downsampling
Interpolation
Soft/hard thresholding
Transform domain hybrid (wavelet/sliding window DCT) filtering (B.Z. Shaick, L. Ridel, L. Yaroslavsky, A hybrid transform method for image denoising, Submitted to EUSIPCO2000, Tampere, Finland, Sept. 5-8, 2000)
Low pass filtering and
downsampling Interpolation
-+ +
Sliding window DCT
domain filtering
Interpolation
+ + +
+ + +
Input
Output
Interpolation
-
+
+
Interpolation
Low pass filtering and
downsampling
Low pass filtering and
downsampling
Interpolation
Sliding window DCT
domain filtering
Interpolation
Sliding window DCT
domain filtering
Wavelet shrinkage, DCT domain sliding window and hybrid processing: image de-noising capability