2.11 _Image Restoration Introduction the ultimate goal of restoration techniques is to improy, - As in image enhancement, ij ibjective proces; i ee . Image enhancement is largely a sul PROCESS, While image in some | pape ss, Thus restoration techniques are oriented towards image restoration is : fing the inverse process in order to recover the Origingy modeling the degradation and app! Contrast stretching is considered an enhancement technique because it is based i it mi the viewer, whereas removal of blur 5 primarily on the pleasing aspects it might present to t e ; Bees deblurring function is considered a restoration technique. 2.11.1 A Model of the Image Degradation/Restoration Process The Fig. 2.11.1 shows, the degradation process is modeled as a degradation function together with additive noise term, operates on an input image f(x, y) to produce a degraded image g(x, y). Given g(x, y). Some knowledge about the degradation function H and some knowledge about the additive noise term (x,y). The objective of restoration is to obtain an estimate to the original input image and in general, the more we know about the H and 1, the closer f (x, y) will be to f(x, y). Where His a linear, position invariant process, then the degraded image is given in the spatial domain by &@&y) = h&y*f&y)+ny) where h (x, y) is the spatial representation of degradation function and * indicates spatial convolution. The convolution in Spatial domain is the multiplication in frequency domain, so we write the model in an equivalent frequency domain as, GW) = Hv) F(uV)+N(@y) ‘where terms in capital letters are Fourier transforms of the corresponding term, fiter(s) FE fxy) Degradation Restoration Fig. 2.11.1 Image degration process Digital Image Processi2.11.2 Noise Models ‘ment and Restoration 2 Principal sources of noise in digital im, =e : me ages arises due to image aquisition and : - - camera, light level; sensor are major factors in producing noise in an image. nae Siepperaixe Images are corrupted during transmission principally due toi E 3 iterference in the channel. e For example, image transmitted during a wireless network mi; be corrupted of lighting or other atmospheric disturbance, ae See 2.11.3 Spatial and Frequency Properties of Noise e Frequency properties refer to the frequency content of noise in the Fourier sense. For example, when the Fourier spectrum of noise is constant, it is called white noise. * Noise is independent of the spatial co-ordinates It is uncorrelated w.r.t. the image itself. (Le. no correlation between pixel values and the values of noise component). 2.11.4 Some Important Noise Probability Density Functions 1) Gaussion Noise : Because of mathematical tractability in both the spatial and frequency domains Gaussian (also called normal) noise models are used frequently in practice. The PDF of a Gaussian random variable Z is given by - 1 -@-p)?/20? Xz = — wy 206 where z represents gray level, j1is the mean of average value of 2 and 6 is its standard deviation. The standard deviation squared, 0° called variance of z. 2) Rayleigh Noise : The PDF of this noise is given by - b 2 -aye )9 forzza P(z)= 4b ; 0 forz O ans B *® < Zmnax and Zxy cannot be an impulse for the same reason that the algorithm outputs the unchanged pixel value, Zxy- By not point, istorton i round in the image Ife Seer then either Zxy = Zmin OF Zxy = Fax: In either case, the value of value and the algorithm output the median value Zed» Which we ‘not a noise impulse. ‘The last step is what the standard median filter median filter replaces every point in the image by the |. This causes cessary loss of detail. point in an reduces salt noise as image. Also it um and minimum ter simply computes the midpoint between the maxim! ed by the filter, Digital Image Processing‘max{g(6,t)} + min | =a AGHESy (DES combines order statistics as well as averaging. Thi, Pike d noise like Gaussian or uniform noise. bee, d Mean Filter delete the $ i ve eo ee tH ging t ae Fay) = aor Y ert) (SE Sxy the value d can range from 0 to mn—1, when d=, it reduces to arithmesi:.. “we choose d= (mn-1)/2, the filter become median filter. - values of d, this filter is useful insituations involving multiple types ‘of such filter is superior to those discussed, Here also we are keeping" ‘image is equal to original image plus noise. Local Noise Reduction Filter ‘Statistical measures of random variable are its mean and ‘are considered as a base of adaptive filters because they are the quatilic’ to the appearance of an image. gives measure of average gray level in the region. ve of noise corrupting f (x, y) to form g (x, y). of the pixels in Sy . =ssitb Proce’ Digital In mee Pree filter should i i retum simply value of g (x, y) noise case in which equal Br oe es tof(x,y) i ry the filter should return a value close to : associated with edges and these should be preser Bes typically yuld be EB ved. . = — a » Thi Saeco ree to one focerallsimge local noise is to be reduced and sit expression for f (x,y) based on these assumptions. — proach to restoration is direct inverse filtering, 2 a , where we compute an transform of the original image simply by dividing the transform of Ny) H(u, v) tells us that even if we image [The inverse Fourier fion whose Fourier transform then the ratio N(u, v)/H (u, y) could easily know the degradation function, we cannot transform of F(u, v)] exactly because js not known. If the degradation has dominate the estimate « to limit the filter frequency (© “around the zero or small value problem n. We know that H(0, 0) is equal to the average value of 3) and that quency domain. e highest value of Hu, v) in the fe (Wiener) oth the degradation function hod is founded on find an estimate f 's minimized. This approach that incorporates by toration process. The me! into the rest es and the objective is to wnare error between them i Disital Image ProcessingImage Enhancement ay, id Regt, writ oy value of the argument. He's assumed that the noise, se the other has zero mean and thatthe gray iy,» it be _|__ Hey) (uy) _ Ca = [ACW +S, 0,95 (u,v) Stay) S-(u) [Hv +8q%) aime (wv)? Guy) HG.v) H(uv)7+Sp (uvySe (9) product of a complex quantity with its conjugate i ~ equal to the magnitude of the complex quantity squared. This result is known as the Wiens filter after N. Wiener who first proposed the concept in the year show 1942. The filter which f consists ofthe terms inside the brackets, also is commonly refered to a5 the minim ns i square error filter or the least square error filter. The Wiener filter does not have the same ‘problem as the inverse filter with zeros in the degradation function unless both H (u, v) and ‘S,(u, v) are zero for the same values (8) of u and v. H(u, v) = Degradation function H*(u,v) = Complex conjugate of H (u, v) |H@w|? = H+@y Ha Gv) $, (u,v) = [N(uv? = Power spectrum of the noise. S_(u,v) = |F(w.v)’ = Power spectrum of the undegraded noise. ‘As before, H(u, y) is the transform of the degradation function and G(u. y) is the of the degraded image. The restored image in the spatial domain is given by ‘Fourier transform of the frequency domain estimate F (u, v). Note that if the no © the noise power spectrum vanishes and the Wiener filter reduces to the inves’ Where we used the fact that the constatt dealing with spectrally white noise, the spectrum|N(u,v)|? is @ considerably. However, the power spectrum of the undegraded ime: oach used frequently when these quantities are not known Of canno! e by the expression. Digital Image Process™®Image Enhancement and Restoration 2.7 Discrete Fourier Transform Image Transforms ‘The one dimensional Fourier transform and its inverse ‘The Fourier transform, F(u) of a si i ti ae } of a single variable, continuous function, f(x) is defined by Fw) = fro feasts where j = =I conversely, given F(u), we can obtain f(x) by means of the inverse Fourier transform. f@) = frye du ‘These two equations comprise the ‘Fourier transform pair’. They indicate the important fact that function can be recovered from its transform. These equations can easily extended to ‘two variable u and v. Fay= ff Fey) PO -dx-dy and similarly for the inverse transform, fey= ff Feye?XOr™. duedy Our interest is in discrete function so we will not dwell on these equations here. ‘The Fourier transform of a descrete function of one variable f(x), x = 0, 1, 2, M~ 1 is given by equation FQ) = M-1 1 =j2mwM for y= LS F¢xyer Pm for u=0, 1,2. Mol M x=0 ee ‘The inverse of this is - M-1 fo= > F(u)-e2™™ for x=0,1,2 M-1 =o In order to compute F(u), we start by substituting values u = 0 in the exponential term summing for all value of x. We then substitute u = 1 in the exponential and repeat the ‘over all values of x. We repeat this process for all M values of u in order to obtain 4 M? summations and multiplications to Fourier transform. It takes approximate Digital Image Processing= Image Enhancement ang Resto, compute the discrete Fourier transform. Like f(x), the transform is a discrete quanti has the same number of components as f(x). The concept of frequency domain follows Euler’s formula. el = cosO+jsin® Y an M-1 FW) = FY £00 feos2munM—jsin2mux/M] x=0 Foru=0,1,2 ...M-1 The values of f (x), in turn are multiplied by sines $ cosines of various frequency, ‘The domain in which values of F(u) range is called frequency domain, because y determines the frequency of components of the transform. Each of the M terms of F(u) is called a frequency component of the transform. A usefull analogy is to compare the Fourier transform to a glass prism. The prism is . physical device that separates light into various color components each depending on jts wavelength content (frequency). The Fourier transform may be defined as ‘Mathematical prism’, that separates a function into various components, also based on frequency content We are consider a light, we talk about its frequency content, similarly, when we consider Fourier transform, we characterize a function by its frequency content. We have seen that components of the Fourier transform are complex quantities. As in the analysis of complex numbers, we find it sometimes convenient to express F(u) in polar co- ordinates. F(u) = [F@e i where Feo = k2 w+? wo)! is called magnitude or spectrum of Fourier transform and 0) = tant (2) called phase angle or phase spectrum. R(u) and I(u) are real and imaginary parts respectively, quantity called power spectrum is given by - P(u) = |F(w/? = Ru)+P?(u) spectral density. ngital Image ProcessingImage Enhancement Example of on DFT Fig, 2.7.1(a) shows . > i 2.7.1(b) shows it spectrum. Both f(s) and —_— an pasueh 1024, A= 1 and K is only 8 points. a a mis centred ext two images depict the same point bul ie ‘The important features to note that, — 1, The height of the spectrum doubled as area under the curve in the x domain doubled. 2. The number of zeros in the spectrum it in the 0 fin Senbied: same interval doubled as the length of the EL ad. — Mpoints—e4 /-———— M points ————= sd © IF (ud) 2K points pois] — M points —+1 (e) (a) Fig. 2.7.1 One dimensional DFT © _ Indiscrete transform, the function f(x) for x= 0, 1,2... (M=1) represents M samples. from its continuous counterpart. These samples are taken at equally spaced but otherwise arbitrary, points. Let Xo denotes first sample and f(xq) is its function. The next sample has taken a fixed interval Ax units away to give £ (x9 +Ax). The ree sample gives us f(x +KAx) and the final sample £(%y+[M-1]Ax) Digital Image Processi)*+¥ prior to computing tt” c m F(0, 0) is located at u s the co-ordinates of Fourier transfor ft this area of frequency domain ieea nsform in a complete the limits of summation are = 1toN. re ofthe transform will be at u= M4, and v= IY, ‘transform at (u, v) = (0, 0) ois 1 MoI Net FOO = > Len X50 yo0 +1 see is the average of f (x, y). In other words, if f (x, y) is an image, the value sform at the origin is equal to the average gray level of the image. Because ; are zero at the origin, F(0, 0) sometimes is called de component of the err in frequency (Fourier) domain. : : ‘smoothing (blurring) is achieved in frequency domain by high frequency that is by low pass filtering. Jow-pass filter that passes all the frequencies within a cirele of radius D, from Fall outside this circle, is called an ideal low pass filter (ILPF 1 if Duy) SD. 0 if D(u,v)2Do and D (u, v) is the distance between a point (u, v) in Digital Image Processing{ Image Processing (PU) 40-10 mR Equation (10.24) reduces to Huy) He geo = [HO 1 Hee lee = Hiv) filter, Since the blurring is usually a low-pass operation high pass filter. ie, This is nothing but the Inverse the Wiener filter in the absence of noise acts as Interpretation of 1 and 2: ‘We have seen that in the presence of only noise, the Wiener filter acts as a low a= filter while in the presence of only blur, the Wiener filter acts as a high pass filter. When both noise and blur are present, the Wiener filter achieves a compromise betwee, a low pass and high pass filter resulting in a band pass filter. The frequency response of the Wiener filter is shown below. Inverse filter epee Fig. 10.8: Wiener filter characteristic 10.6.1 Drawback of Wiener Filters : Though the Wiener filter offers an optimal compromise between low pass and high pass ‘Operations it is plagued by four problems, 1, If you remember, during the derivation of the wiener filter, we minimized the mean squared error (m.s. ©) by using the orthogonality condition. The m.s.c. criteria is 00" very effective when images are restored for the human eye. The reason being that the m.s.¢, weighs all errors equally regardless of their location in the image, The human ey’ more than what the human eye would prefer.f(x, y) > original image before degradation g(x,y) > observed degraded image (x, y) > additive noise y) — Spatially invariant point spread function (Degradation function) ) — Restored image 1 is divided into two parts viz, the degradation block and the restoration on the degradation block first. The linear image model is given by See ff £(K,)-h(x-k, y-) dk di +n (x, y) (10.1) »y) = f(%y)*h@y) +n y) (102) original image gets convolved with the degradation function h (%, y) degradation function (convolution) we need to apply inves degraded image. n (10.2) is given by the formula, Dh&-k,y-D Fk, D+a%,y) (0g also know * (x, y), of a few degradation functions Point Spread Functions (P.S.F)PEvilenc i Image Re * (tis Gaussian Bal testoration © ako?) Shaped). Fig, 10.3 image acquisition or C.C.D interaction es, the blur function is separable ic. h(x,y) = h@&)-hO) ise what has been stated so far. ge, f (x, y), gets degraded by a deg ‘added with noise n (x, y)- This relation is 10.4) radation function (bluring function), given by the filtering operation f(xy) thy) ta y) (10.5) g(x,y) = blurring functions 0 discussed some of the Degradation Model : ize that the degradation (blurring) process is nothing but ce of noise, we reall ation i.¢., gy) = f(x,y)*h «wy(107) finding the solution is a ming is by applying inverse filtering wolved by a blurring function h (x, y) and g (x, y), we take the inverse of h (x. )). n (10.3) can be written as, REN -k yD EK) (10.8) (u,v) xF (u,v) Guy) Huy)