UNIVERSITY OF ENGINEERING & TECHNOLOGY, MARDAN

IMAGE RESTORATION

BASIC IDEA

MODEL THE DEGRADATION & APPLYING THE INVERSE PROCESS IN ORDER TO RECOVER THE ORIGINAL IMAGE SOME RESTORATION TECHNIQUES ARE BETTER FORMULATED IN SPATIAL DOMAIN, WHILE OTHERS ARE BETTER SUITED FOR FREQUENCY DOMAIN

SPATIAL PROCESSING IS APPLICABLE WHEN THE ONLY DEGRADATION IS ADDITIVE NOISE FREQUENCY DOMAIN PROCESSING EASIER TO IMPLEMENT TO DEAL WITH IMAGE BLUR

MODEL OF DEGRADATION/ RESTORATION PROCESS

PROCESS
GIVEN g(x,y)- A DEGRADED IMAGE, SOME KNOWLEDGE ABOUT DEGRADATION FUNCTION H, & SOME KNOWLEDGE ABOUT ADDITIVE NOISE TERM (x,y), OBJECTIVE OF RESTORATION IS TO OBTAIN AN ESTIMATE OF THE ORIGINAL IMAGE
THE MORE WE KNOW ABOUT H AND , THE CLOSER IS OUR ESTIMATE TO f(x,y)

SPATIAL DOMAIN: g(x,y) = h(x,y)*f(x,y)+(x,y) FREQUENCY DOMAIN: G(u,v)=H(u,v).F(u,v)+N(u,v)

IMAGE DEGRADATION/ RESTORATION

WHEN WE CONSIDER ONLY NOISE DEGRADATIONS, WE WILL TAKE H AS IDENTITY OPERATOR

NOISE MODELS

PRINCIPAL SOURCES OF NOISE IN DIGITAL IMAGES

DURING IMAGE ACQUISITION AND/OR DURING IMAGE TRANSMISSION

PERFORMANCE OF IMAGING SENSORS IS AFFECTED BY ENVIRONMENTAL CONDITIONS DURING ACQUISITION & QUALITY OF SENSORS DURING TRANSMISSION, IMAGES ARE DISRUPTED DUE TO CHANNEL INTERFERENCE

SPATIAL & FREQUENCY PROPERTIES OF NOISE

PARAMETERS THAT DEFINE SPATIAL CHARACTERISTICS OF NOISE CORRELATION OF NOISE WITH IMAGE FREQUENCY CONTENT OF NOISE

WHEN FOURIER SPECTRUM OF NOISE IS CONSTANT, IT IS WHITE NOISE

FOR NOW, ASSUMING THAT NOISE IS INDEPENDENT OF SPATIAL COORDINATES ASSUMING THAT THERE IS NO CORRELATION BETWEEN NOISE COMPONENTS & PIXEL VALUES

SOME IMPORTANT NOISE PDFs

GAUSSIAN NOISE RAYLEIGH NOISE ERLANG (GAMMA) NOISE EXPONENTIAL NOISE UNIFORM NOISE IMPULSE (SALT & PEPPER) NOISE

NOISE PDFs

NOISE PDFs

EQUATIONS AS PER TEXTBOOK THE PDFs PROVIDE USEFUL TOOL FOR MODELLING A BROAD RANGE OF NOISE CORRUPTION SITUATIONS

GAUSSIAN NOISE - ELECTRONIC CIRCUIT NOISE & SENSOR NOISE RAYLEIGH NOISE - RANGE IMAGING EXPONENTIAL & GAMMA LASER IMAGING IMPULSE NOISE QUICK TRANSIENTS (FAULTY SWITCHING) DURING IMAGING

NOISE PDFs
SPAN FROM BLACK TO NEAR WHITE IN 3 INCREMENTS NOISES OF ALL 6 TYPES WILL BE ADDED TO THIS TEST PATTERN

NOISE PDFs

NOISE PDFs

NOISE PDFs

PARAMETERS OF NOISE WERE CHOSEN SO THAT TEH HISTOGRAMS FOR 3 SECTIONS OF TEST PATTERN WOULD START TO MERGE CLOSE CORRESPONDENCE BETWEEN HISTOGRAMS & PDFs OF NOISE DIFFICULT TO DIFFERENTIATE VISUALLY BETWEEN THE FIRST 5 IMAGES, EVEN THOUGH THEIR HISTOGRAMS ARE DIFFERENT IMPULSE NOISE GIVES SALT & PEPPER APPEARANCE & VISUAL INDICATION OF TYPE OF NOISE CAN BE OBTAINED

PERIODIC NOISE

ARISES TYPICALLY FROM ELECTRICAL OR ELECTROMECHANICAL INTERFERENCE DURING IMAGE ACQUISITION ONLY TYPE OF SPATIALLY DEPENDENT NOISE WE WILL STUDY (FOR OTHERS, SPATIAL INDEPENDENCE WAS ASSUMED EARLIER) CAN BE REDUCED CONSIDERABLY USING FREQUENCY DOMAIN FILTERING CONSIDER THE IMAGE IN FIG 5.5 a

SEVERELY CORRUPTED BY (SPATIAL) SINUSOIDAL NOISE OF VARIOUS FREQUENCIES

PERIODIC NOISE
FOURIER TRANSFORM OF A PURE SINUSOID
PAIR OF CONJUGATE IMPULSES LOCATED AT THE CONJUGATE FREQUENCIES OF THE SINE WAVE

IMPULSES IN SHAPE OF CIRCLE

ESTIMATION OF NOISE PARAMETERS

PERIODIC NOISE

PARAMETERS ESTIMATED BY INSPECTION OF FOURIER SPECTRUM OF IMAGE

PERIODIC NOISE TENDS TO PRODUCE FREQUENCY SPIKES THAT CAN BE VISUALLY DETECTED AS WELL

ATTEMPT TO INFER THE PERIODICITY OF NOISE COMPONENTS DIRECTLY FROM THE IMAGE (POSSIBLE ONLY IN SIMPLE CASES) AUTOMATED ANALYSIS POSSIBLE WHEN NOISE SPIKES ARE PRONOUNCED OR WE HAVE SOME KNOWLEDGE ABOUT THE LOCATION OF FREQUENCY COMPONENTS OF INTERFERENCE

ESTIMATION OF NOISE PARAMETERS

PARAMETERS OF NOISE PDFs MAY BE KNOWN PARTIALLY FROM SENSOR SPECIFICATIONS BUT USUALLY ESTIMATION IS REQUIRED FOR A PARTICULAR IMAGING ARRANGEMENT IF IMAGING SYSTEM IS AVAILABLE

CAPTURE A SET OF IMAGES OF FLAT ENVIRONMENTS

OPTICAL SENSOR: IMAGING A SOLID GRAY BOARD WHICH IS UNIFORMLY ILLUMINATED

IF IMAGES GENERATED BY SENSOR ARE AVAILABLE

POSSIBLE TO ESTIMATE PDF PARAMETERS BY SMALL PATCHES OF ALMOST CONSTANT

ESTIMATION OF NOISE PARAMETERS

VERTICAL STRIPS OF 150x20 PIXELS HISTOGRAMS BASED ON IMAGE DATA FROM STRIPS SIMILARITY IN HISTOGRAM SHAPES VISIBLE

ESTIMATION OF NOISE PARAMETERS

USE DATA FROM IMAGE STRIPS

CALCULATE MEAN & VARIANCE OF GRAY LEVELS STRIP DENOTED BY S, TAKEN AS SUBIMAGE

ZiS ARE THE GRAY LEVEL VALUES OF PIXELS IN S AND p(zi) ARE CORRESPONDING NORMALISED HISTOGRAM VALUES SHAPE OF HISTOGRAM IDENTIFIES CLOSEST PDF

RESTORATION IN THE PRESENCE OF NOISE ONLY-SPATIAL FILTERING

ONLY DEGRADATION IS NOISE

g(x,y)=f(x,y)+(x,y) G(u,v)=F(u,v)+N(u,v)

If noise is periodic, it is possible to estimate N(u,v) from spectrum of G(u,v). Then F(u,v)=G(u,v)-N(u,v) Spatial Filtering is the choice when only additive noise is present Mean Filters

Arithmetic Mean Filter Geometric Mean Filter Harmonic Mean Filter Contraharmonic Mean Filter

Illustration of Mean Filters

Illustration of contraharmonic filters

Suitability of filters
Arithmetic and geometric mean filters are well suited for random noise like Gaussian or uniform noise Contraharmonic filter is well suited for impulse noise but it MUST be known whether the noise is dark or light to select the proper sign for Q
The results with choosing the wrong sign for Q are disastrous (Fig 5.9)

Using wrong Contraharmonic filters

ORDER STATISTICS FILTERS

NON-LINEAR SPATIAL FILTERS, WHOSE RESPONSE IS BASED ON ORDERING (RANKING) THE PIXELS CONTAINED IN TEH IMAGE AREA ENCOMPASSED BY THE FILTER AND THEN REPLACING THE VALUE OF CENTRAL PIXEL WITH THE VALUE DETERMINED BY THE RANKING RESULT

MEDIAN FILTER MAX/MIN FILTER MIDPOINT FILTER ALPHA-TRIMMED MEAN FILTER

MEDIAN FILTER
BEST KNOWN ORDER-STATISTICS FILTER
FOR CERTAIN TYPES OF RANDOM NOISE, THEY PROVIDE EXCELLENT NOISE REDUCTION, WITH CONSIDERABLY LESS BLURRING THAN AVERAGING FILTERS PARTICULARLY EFFECTIVE IN PRESENCE OF BOTH BIPOLAR & UNIPOLAR IMPULSE NOISE

MEDIAN FILTER

MAX/MIN FILTERS

COMPARISON OF FILTERS

ADAPTIVE FILTERING

ADAPTIVE MEDIAN FILTERING

PERIODIC NOISE REDUCTION BY FREQUENCY DOMAIN FILTERING

BAND-PASS FILTERS

NOTCH FILTERS

BAND REJECT FILTERS

BUTTERWORTH BAND-REJECT FILTERING

BAND-PASS FILTERS

NOTCH FILTERS

NOTCH FILTERS

ESTIMATING THE DEGRADING FUNCTION

EXPERIMENTATION
If equipment similar to that used to acquire degraded image is available

MATHEMATICAL MODELLING
Example, physical properties of atmospheric turbulence

MATLAB COMMANDS
imhistmatch histeq stretchlim