IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING,

VOL. ASSP-27, NO.

6, DECEMBER 1979

62 5

Image Enhancement by Stochastic Homomorphic Filtering

ROBERT W. FRIES AND JAMES

w.MODESTINO, MEMBER, IEEE

D = log,, l / T

Abstract-The problem of image enhancement by nonlinear twodimensional (2-D) homomorphicfiltering is approached usingstochastic models of the signal and degradations. Homomorphic filtering has been previously used for image enhancement, but the linear filtering operation has generally beenchosen heuristically. In this paper stochastic image models previouslydescribedandanalyzed by the authors are used to model the true image and interfering components (shadowsand salt-and-pepper noise). The problem of designing the linear filter can then be formulated as one of linear least mean-squared error (wiener) filtering. Examples of processing of typical real-world images are included to indicate the obtainable results.

(1)

I. INTRODUCTION PROBLEM in representing data in any format is emphasizing the significant features while minimizing distracting detail. Solutions to this problem when the data consists of images are called imageenhancement techniques. Elements in images which distract the viewer from relevant information are shadowsandsalt-and-pepper noise. This is the case for eitherphotochemicalorphotoelectronic image sensing and display systems. For example, shadows frequently obscuredetail in ordinaryphotographs; evenwhen detail is present within the area darkened by the shadow, it is often subdued because the contrast in the darkened area is less than that in the surrounding brightregions.' The same effect occurs in therecordinganddisplay .of images by photoelectronic means although here the problem is further aggravated by the limited dynamic range generally available.Many TV monitors and storage CRT displays, for example, are not capable of more than 32 distinct gray levels. The severity of this restriction is best appreciated in comparison to the capabilities afforded by a good quality photograph under subjective humanevaluation. According to Zoethout [ 1 ] , the ratio of the smallest variation in intensity discernible by the eye to the overall intensity is generally considered to be 1/100. Using thestandarddefinition [2] of theoptical density D of a photographic emulsion
Manuscript received March 28, 1979; revised July 23,1979.This workwassupportedinpartby the Office of NavalResearchunder Contract N 0 0 0 1 4 - 7 5 4 0281. R. W. Fries was with the Department of Electrical and Systems Engineering, Rensselaer Polytechnic Institute, Troy, NY 12181. He is nowwith the Pattern Analysis and Recognition Corporation, Rome, NY. J. W. Modestino is with the Department of ElectIical and Systems Engineering, RensselaerPolytechnic Institute, Troy, 12181. NY n photography _ of 'This is due to the common practice i modern _ _ exposing film so thatshadows fall on the toe of the D-log E curve (Cf. [19], pp. 29-32).

where T is the transmittance of the emulsion, we conclude that the smallest discernible change in density is roughly log, lOl/lOO = 0.0043.Typicalphotographicemulsions (cf. [2, p. 691) cover a range of densities from 0 to 1.5. Hence, assuming grain noise is negligible, there are approximately 347 discernible density levels in a good quality black-and-whte photograph, far exceedingthe capabilities of many photoelectronic displays. The mechanism by which the eye achieves its wide dynamic range suggests techniques which dynamic by the range of photographicmaterialsand electronic displaysmaybe extended. The sensitivity of the eye to local changes is described by Hochberg [3]. Within a broad range of illuminations, the apparent brightness of a region is a function of the ratio between its intensity and that of its surrounding region and not of itsabsoluteintensity alone. It is as if theeyetakes the logarithmoftheincoming light intensities and then passes these signals through a spatial high-pass filter which attenuates the dc component. This contention is supported by Stockham [4] who carries it to the point of estimatingthefrequency response of the high-pass filter. Additional work along these lines has been described by Mannos and Sakrison [20] in the context of image coding. One method, then, of extending the dynamic range of recorded images is to take the logarithm of the original intensities, attenuate slowlyvarying components which contribute little information to the eye (and, indeed act as a source of degradation) while simultaneously enhancing rapidly varying components to which the eye is most sensitive, and finally, taking the exponential of the processed image. Photographers have, to a degree, accomplished this processing in recording images on filmby using afast-acting developer withlittle agitation;exhausted chemicals from heavily exposed areas diffuse into adjacent lightly exposed areas causing them to develop even less andtherebyemphasizingthe transitions from light to dark [2]. For images stored in digital form, this processing is readily accomplished by adigital computer. This paper describes a specific digital implementation of a scheme for enhancingrapid local variations due to object boundaries and deemphasizing more gradual changessuchas those which occur as a result of shadows while simultaneously controlling the degrading effects due to salt-and-pepper noise. In order to formulate this problem the observed image intensity so(x) will be considered to be the product three compoof nents: the true reflectivity of the objects in the image ~(x), the
'

0096-3518/79/1200-0625$00.75 0 1979 IEEE

626

IEEETRANSACTIONS ON ACOUSTICS,SPEECH, AND SIGNAL PROCESSING,VOL. ASSP-27, NO. 6, DECEMBER 1979

nonconstant illumination function component n (x). That is,

so (x) = r(x) . i(x) . n (x).

i(x),and

a white noise (2)

filter is provided in Section VI. Some typical results are illustrated in Section VI1 whichis followed by a summary and conclusions in Section VIII. We begin with some preliminary discussions on random fields as image models.
11. PRELIMINARY DISCUSSION Theapproach described earlier is not completeuntilthe stochastic descriptions of f,(x), fi(x), and &(x) are specified. The models to be considered here will be a family of random variables cfx(o), E R 2 } defined on the plane where o dex notes the dependence on a fxed probability space (a, Q ,Pj. Wong [lo] refers to such a collection as a random field, but notes that it is also called a stochastic process with a severaldimensional time. In the discussion that follows, for convenience, the dependence upon the underlying probability space will not be explicitly denoted;f(x) will be written for &(a). Furthermore, consideration will be restricted to zeromean fields of a second-order (variances exist). The secondorder properties are described by the covariance function of the random field which is given by

We notethatthe noise here ismodeledas a multiplicative process in the intensitydomain or, equivalently, additive in the density domain. There is substantial evidence to justify this for either photochemical or photoelectronic sensing and display systems.The discussion in Andrews andHunt[21] is particularly illuminating in this regard. In general, however, the noise process n(x) will depend upon the signal in a rather complicatedfashion. We have chosen to disregard this dependence and assume a signal-independent noise process n ( ~ ) . Again this is common practice in digital image processing (cf. [21, pp.20-231). Taking thelogarithmof both sidesof (2) results in the density image

g(x)

e In so (x) = f,(x) + fi(x) +f n (x)

(3)

where f,(x) & In r(x), fi(x) = In i(x), and finally &(x) = In n(x). Ifprecise stochastic descriptions of the signalsf,(x), fi(x), and f,(x) areavailable, the problem of extracting the true image fromtheilluminationand noise can be formulated as one ofWiener or linear least mean-square filtering, where the signal is f,(x) and the noise isfi(x) +fn(x). Passing the result of this filtering through an exponentiation operation yields an estimate ?(x) of the true image as represented by the reflectance process r(xj. A block diagram of theindicated processing is provided in Fig. 1. Thetechnique of sandwiching a linear operationbetween two complimentary nonlinear operations (homomorphic filtering) [7] as applied to image enhancement is not new; what is unique about the approach described here is the use of explicit stochastic image models as a guide in the choice of a linear-filter transfer function. In particular, we develop explicitstochasticmodelsfor boththe reflectance process f,(x) and the illumination process fi(x). The reflectance process f,(x) has been developed to exhibit predominant and pronounced edge structure, such as might be expected of realworld imagery. This process possesses considerably high spatial frequency components. The illumination processfi(x), on the other hand, has been specifically developed to possess relatively largelow spatial frequency components typical of illumination fields. These stochastic models are described in Sections 1 1and IV, 1 respectively. The noise processf,(x) is described in Section V while a description of the digital implementation of the Wiener

Rff(X,.Y) = E Cf(x>f(u>); x,.Y E R 2

(4)

where E { .} represents expectation the operator over the probability space (a,Q , P). All of the random fields to be presented here have been constructed so as to possess a covariance function invariant under all Euclidean motions; according to the definitions ofWong [lo] , they are homogeneous and isotropic. For these fields, (4) can be written E{f(x+u)f(x)}=Rff(IIu

11);

uT=(u1,u2)ER2

(5)

where I(u 11 represents the Euclidean norm defined in terms of an inner product according to 11 u 11' = ( u , u ) = u; + uz. Since the random fields are homogeneous, their power-spectral density is given by S f f ( o )=

[ R f f ( u )exp {-i
R2

(w, u ) } du

(6)

where o = (a1, 2 ) represents a spatial frequency vector w and du is the differential volume element in R 2 . Furthermore, for isotropic fields, R f f ( u )= Rff(llu 11) and (6) can be evaluated up to the functional form with the aid of a theorem of Bochner [ 11] yielding
Sff(0) = Sff(LQ)= 27r

L-

h R f f ( h )Jo(hS2) d h

(7)

where S2 4 11 o 11 = (a; &)'I2 t

represents radial frequency

FRIES AND MODESTINO: IMAGE ENHANCEMENT BY STOCHASTIC HOMOMORPHIC FILTERING

621

and Jo( - ) denotes the ordinary Bessel function of the first kind of order zero. Thus, S f f ( * ) and R f f ( are related through a Hankel transform[121 , [ 131 . The enhancement problem aswe haveposed it consists of emphasizing fr(x), while minimizing the degrading effects of &(x) +f,(x); specifically, wewill lookforthe linear least mean-squaredestimateof f,.(x) based on observing g(x). If we haveknowledge ofthepower-spectral densities2 Srr(o), Sii(o), and & ( ) associatedwith fr(x), &(x), and f,(x), ,a respectively, and these three fields are uncorrelated, the twodimensional linear least mean-square filter possesses the transfer function
e )

Fig. 2. Rotation of Cartesian coordinate axes.

the probability density function p ( t ) of the distance between events begiven by the exponential distribution function p ( t ) = XrFArt;

t20 .

(1 1)

which is simply an extension to two dimensions of the Wiener filter [SI. We will now proceed to describe the randomfields which are to be used as the reflectance, illuminance, and noise models, respectively. 111. REFLECTANCEPROCESS The process to be considered as a model of the reflectance will be referred to as the rectangular process. This process divides theplane into regions whose boundariesform rectangles, and then assigns intensities to these regions such that theintensitywithineach region is constantand possesses specified correlation properties with intensities in contiguous regions. It was originally developed to satisfy the need for an edge model for the design of a linear filtering operation to be used in edge detection [15] . Roughly speaking, it is a twodimensional extension of a generalization the of random telegraph wave [ 141 . As a first step in describingthis process, themethod of partitioning the plane into rectangles will be detailed. A fundamental role in this stochastic image model will be played by the vector-valued random field N(x) which provides a twodmensional generalization of a counting process. In particular, suppose the vector x is obtained from x according to ?=Ax where A is the matrix

, The quantity X represents the average number of events per unit distance. Consider now the random field {&(x), x E R 2 } ,which undergoes transitions at the boundaries of the elementary rectangles defined by {N(x), x ER2}. The gray levelassumed throughout any elementary rectangle will be assigned so as to be zero-mean Gaussian3 with variance u, and correlated with the gray levels in contiguous rectangles. More specifically, assume thattherandomorientation 0 E [-n, hasbeen 771 chosen and that k transitions have occurred between the two points x and x + u. It will be assumedthat
E(f(x+u)f(x)Ie,k}=u,2pyk; k = 0 , 1 ; * * ,
(12)

where I p r I < 1. The quantity pr is the correlation coefficient governing the spatial evolution the of random amplitude process. Forexample, let Xi,j representtheamplitude,or gray level, assigned after i transitions in the Tl direction and j transitions in the direction. The sequence {Xi,j} can then be generatedrecursively according to

z2

x.. = p r x.1 , j a,] z-

+ P r x i , j - l - dxi-1,j-I + W , j

(13)

A=

[-

sin e cOse

cos e

.defined for some 0 E [-n, . The transformation (9) then n] results in a rotation of the plane in terms of the Cartesiancoordinate axes (xl, xz), as illustrated in Fig. 2. Consider now the vector-valued randomfield defined by

where { W i , j } is atwo-dimensionalsequenceofindependent and identically distributed (i.i.d.) zero-mean Gaussian variates withcommon variance u$ = uz(1 - p:) Thesequence defined by (13) has an alternative interpretation as the output of a recursive two-dimensional digital filter excited by a white noise field. It can be shown that E
{ x i , j x i + k, ,j + k ,

} = 0,p r

Ik, I+ Ik, I

(1 4)

so that the condition expressedby (1 2) is indeed satisfied. Typical computer-generated realizations of the resulting random field are illustrated in Fig. 3 for selected values of p r and X Note that X, controls the edge density, while p r indi, . cates the degree of correlation of the intensities in adjacent where 0 is uniformly distributed on [-n, n] and {Ni(t), regions.The displayed images hereandthroughoutthe ret > 0 , i = 1, 2 are mutually independent one-dimensional } mainder of this paper are square arrays consisting of 256 elecounting processes, i.e., Ni(t) represents the number of events , which have occurred in the interval [0, t ]. Specifically, let ments, or samples, on a side. Where a value of X is specified, it is measured in normalized units of events per sample distance so that-t&re are, on average, 256Xr transitions along In general, the power-spectral density of a random fieldf(x) will be each of the orthogonalaxes. denoted by Sf-a). In situations wheresubscripts a-e-used, suchas
3The Gaussian assumption here isnot critical and is easily removed.

fr(x), the corresponding power-spectral density will be distinguished by the associated subscript, i.e., S,,(O) in the case offr(x).

628

IEEE TRANSACTIONSACOUSTICS, ON SPEECH,

AND SIGNAL PROCESSING,

VOL. ASSP-27, NO. 6, DECEMBER 1979

Fig. 3. Selected realizations of the reflectance process. (a) h, = 0.0125, P, = -0.9. (b) A= 0.0125, P, = 0.0. (c) hr= , 0.0125, p , = 0.5. (d) A, = 0.025, p, = -0.9. (e) h, = 0.025, P, = 0.0. (f) h, = 0.025, p, = 0.5. (8) A 0.05, p r = -0.9. , (h) A, = 0.05, p, = 0.0 (i) h, = 0.05, P, = 0.5.

In [I51 it has been shown that the power-spectral density of this process is given by

hand,thetransition across anedge boundary is much more gradual. These properties areeasily related to any particular real-world image. IV. ILLUMINATION PROCESS

I 2 . (15)

Although realizations of this process may not resemble any particular real-world image, the general properties of regions of constant reflectivity and boundaries of regions being parallel arecommon to many real-world images. More specifically, we do not claim that the model proposed here is universally representative of real-world imagery except in a certain qualitative sense. This model is completely defined, up to a scale factor, by the two parameters X, and p,. Theparameter X, represents the edge busyness associated with an image, while p , is indicative of the degree ofabruptness across an edge boundary.For p r large (in magnitude) and negative, there is an abruptalmostblack-to-whiteorwhite-to-black transition across an edge boundary. For p > 0, on the other

In the previous section a two-dimensional random field which possesses an inherent rectangular structure was described and analyzed. While this random field is a reasonable model of objects, shadows do not generally exhibit this rectangular mosaic, but rather a muchmorerandom edge orientation. Here the construction and properties of a class of two-dimensional random fields, which provide a more appropriate model illumination for in the sense that individual realizations do not exhibit this rectangular pattern will,be presented. Consider the random partition of the plane by a seriesofsensed (directed) lines whichcan be described as a marked-point process. Specifically, anarbitrary sensed line can be described in terms of a 3-tuple (r, 0 , d ) where r repre-

FRIES AND MODESTINO: IMAGE ENHANCEMENT

\

BY STOCHASTIC FILTERING HOMOMORPHIC

629

regions tend to be surrounded by light regions, while for p i = -0.9 there is a tendency for light regions to be surrounded by dark regions and vice versa. Indeed, in the latter case there is almost a black-to-white or white-to-black inversion across an edge boundary. An expression for the autocorrelation function given by (4) can be obtained, as in [ 1 7 ] ,with the result

Fig. 4. Parameterization of a directed line segment.

sents the perpendicular distance, or radius, to the line in question, 6 E [-n, represents the orientation of the vector n] of length r drawnfromthe origin andperpendicular to this line, and d E {- 1 , l } specifies the sense of the line where Fig. 4 illustrates a case of d = 1 (pointingcounterclockwise). As shown in Fig. 4, by virtue of the direction imposed on this line segment, the planeis partitioned into two distinct regions, R (right of line) and L (left of line), such that R U L = R 2 . Now consider the set of lines represented by the sequence {(ri, B i , di)}. Here {ri} representstheeventpositions associated withahomogeneous Poisson process {N(r),v> O } , with intensity A eventslunit distanceevolving according to the radial parameter Y. Thesequence {ei} represents an(i.i.d.) sequence of random variables uniformly distributed on [-x, , n] while the sequence Idi} is likewise i.i.d., assuming the binary values +1 with equal probability. Finally, the sequences {ri}, ( 0 3 , and {d;} are assumed to be independent. Now notethatthe set of lines generatedbytheprocess described abovewilldivide the plane into disjoint polygonal regions. For any point x, define NLR (x) to be the number of left-to-right crossings of sensed lines incurred in moving along a straight path from the origin to the point x; similarly, define NRL(X) be the number of right-to-left crossings.Observe to that N L R ( x ~ )N L R ( x ~and NRL(x1) = N R ~ ( x Z ) any = ) for twopoints x1 and x2 which are the same region. Assign a label k to each region as k =NLR(x)- NRL(x) where x is a point in the region. This label can in turn be used to assign a gray level to each region by using it to index into a sequence {X,} of zero-mean Gaussian random variables. In particular, assume that the sequence {X,} is generated recursively according to xk=pix,-, and X,=&&+,

(18)
where I,( is the modified Bessel function of the first kind of order k. Similarly, thecorrespondingpower-spectraldensity evaluated accordingto (7) is given by
e)

which is illustrated inFig. 6 for various values of pi. One noteable characteristic of this random field is that for a small compared to Ai, the power-spectral density behaves approximately as i.e., Sii(a) a has singularity at the origin except for p i = - 1. This high concentration of energy at low spatial frequencies is a direct result of the construction procedurewhich allows relatively large correlationbetween gray levels in regions relatively far apart. We feel that this characteristic is typical of random illumination processes and as a result it was purposely built into the construction procedure. The effect on a real-world image of multiplying the intensity by the exponential of a realization of the polygonal processis illustrated in Fig. 7 . Note how the illumination process seems simply to have added extra shadows to the building. Although somewhat of an exaggeration, this figure has been included to illustrate the nature of effects which may be caused by the proposed random illuminationprocess.

+w,; +w,;

k=1,2,..
(1 6)

k=-l,-2;..,

with {W,} an i.i.d. zero-mean Gaussian sequence with variance u$ = ( 1 - p f ) u ; , andthe initial value X , chosen to have zero-mean Gaussian distribution with variance u;. Note that this sequence is stationary with the covariance given by

Selected computer-generated realizations of the resulting random field are illustrated in Fig. 5 for various values of p i and Xi h/n. The quantity Xi canbe shown to representthe average edge density along any randomly chosen line segment on the pline. From Fig. 5 observe that, for pi = 0.5, light

V. WHITE NOISE PROCESS Inanyphotographic image that is scannedthere issome noise present that is independent from pixel-to-pixel; due to its appearance it is often referred to as salt-and-pepper noise. Here we will model this noise, which is represented by f,(x), as white Gaussian noise. While the Gaussian assumption is not critical to the analysis which follows, it is interesting to note the observation of Selwyn (cf. [2, p. 1251) on noise due to film grains. He notes that if the spot areacovered bya beam used to measure the density of a photographic emulsion is much larger than the projection of individualgrains then the density distribution will be Gaussian if the grain noise is small, and presumably homogeneous. fact, following In the relationship between the size of the scanning aperture and variance of the measured densities has been found to hold

where u i is the variance of the density, G is a measure of the

630

IEEE TRANSACTIONSACOUSTICS, ON SPEECH,

AND SIGNAL PROCESSING,

VOL. ASSP-27, NO. 6 , DECEMBER 1979

Fig. 5. Selected realizations of the illumination process. (a) hi= 0.0125, p i = -0.9. @) hi= 0.0125, p i = 0.0. (c) h i = 0.0125, p i = 0.5. (d) hi = 0.025, p i = -0.9. (e) hi = 0.025, p i = 0.0. (f) hi = 0.025, pi = 0.5. (g) hi = 0.05, pi = -0.9. (h) hi = 0.05, p i = 0.0 (i) hi = 0.05, p i = 0.5.
m

=; ;

IO

Normalized Spatial Frequency W 2 7 X i

Fig. 6. Power-spectral density of the illumination process.

granularity of thefilm, and a is the area of the scanning aperture. If the scanning aperture is considered to be a spatial low-pass filter whose bandwidth is proportional to 1 / 6 , then (20) implies that the noise power present in the density measurement is proportional to the low-pass filter bandwidth squared. Since this is the effect that would be observed if

the grainnoise were white,this modeling assumption is justified. The power-spectral density of this noise can be expressed then as
~ r Z n ( W 0; =

(2 1)

where 0; is the varianceof the samples.

FRIES AND MODESTINO: IMAGE ENHANCEMENT

BY STOCHASTIC HOMOMORPHIC

FILTERING

631

of illumination and noise powersto reflectance power, respectively. These parameters provide a measure of the degree of degradation caused by shadows salt-and-pepper and noise. More specifically,we have

while

so that the filter is completely specified terms of the quantiin ties y i , -yn p r , h,., pi, and hi. These quantities can be defined, at least empirically, for wide range of images. Finally, a observe that, for the given power-spectral densities, the filter is radially symmetric, i.e.,it is a function onlyof s2 = Ilwll. Since the final objective of this process is to make the best use of the limited dynamic range available for display, it wiU prove expedient to include a gain factor A . chosen such that the power leaving the linear processor equals that entering. It follows then that do is given by
(25)

where Sgg(o)is the power-spectral density of the fiiter input g(x) in (2 j and is given by

The description of thefilterprovided above is for a continuous spatial domain; we have access only to sampled data. Hence we seek a (2 - D) digitalfilterwithsystemtransfer function H o ( z l ,z2) whose frequency response4 approximates Ko(!2). Specifically, the Wiener filter has been implemented as a two-dimensionalinfinite impulse response (IIR) [lS] digitalfilter whose point spread functionexhibitsquadrant symmetry. The choice of an IIR filter was based on the computationaleconomies which resultfromtheability to implement the filter with a recursive structure. Consider first a filter with thefolrowing recursive structure:

where5 D 2 { ( i , j ) :i = O , I ; * . , N , j = O , I ; * - , M } a n d D ' = D - {O 0 ) j . At this point we will assume that the coefficients (, of y in (32) have been chosen such that the resulting filter is (C ) ObFig. 7. Effects of the illumination process. (a) Original. (b) Iflumina- stable when recursing from the upper left-hand corner. serve that the resulting filter will have a nonzero response only tion. (c) Original with illumination. in the lower right quadrant. Since the filter described in (22) exhibits radial symmetry, it is desirable for the resulting VI. LINEAR FILTERDESCRIPTION digital filter to exhibit four-quadrant symmetry. AND IMPLEMENTATION One method for achieving a point spread function with this Using themodelspresentedabove,thelinearleastmeaninherentsymmetry is to allow repeatedapplication of the squarederror (Wiener) filterforestimatingthereflectance process will now be derived. The systemtransferfunction 4The frequency response of the fwo-dimensional digital f'iiter is simgiven by (8) can be rewritten as
i ply H o ( z , , 22) evaluated for z = elwi, i = 1, 2. We assume the spatial in units of radians sample per frequency variable wi is measured distance. 'The support D of the recursive two-dimensional filters taken here will be such as to force a quarterplane causality condition. That is, D consists of a finite number of points above and to the left of the pixel i position (i, j ) . n

which suggests defining the parameters yi and -yn as the ratios

632

IEEE TRANSACTIONS ON ACOUSTICS, SPEECH,

AND SIGNAL PROCESSING, VOL. ASSP-27, NO. 6, DECEMBER 1979

Fig. S. Point spread function of the composite filter.

TABLE I TYPICAL FILTER PARAMETERS dB FOR y n = -20

Shadow
Yi(dB)
hi Pi

hr

Dr

Characteristic Soft
Medium

-6
-3

.0125
.I3125

0.5

.05

0.5

.05

0
@

'

-.e650

,7833 -.a240

-.1957 ,1831 -.@147 .2376

.2681 -.2216 .2072 -.0160

-.e503

.771.3

-.2326

Harsh

.0125

-.9

.@5

-.a817

,7959 -.0118 .5109 -.4?18

1 . 1

.41;58 -.0065

line 45" to the axis, i.e., h ; , = hj,i. Similar properties extend, ~ of course, to the composite filter represented by H(zl, z 2 ) . A computer program has been written for determining the coefficients boo,bol , b l l ,bo2,aol ,a l l , and aO2according to an iterative gradient procedure to result in a frequency response for H ( z l , z 2 ) which provides a least mean-squared error approximation to the desired response Ho(i2), The details of ~ ( z 1 ~ z 2 ~ ~ H 0 ( z 1 ~ z 2 ) t H 0 ( z ~ ' ~ z 2 ~ t H 0 ~ z Z , ~ z ~ 1 ~this program are described in [6]. In Table I wesummarize z;') t Ho(z;l, (28) the results of t h s procedure for selected values of yi, Xi,pi, A,, and p r , all with yn = -20 dB. Here we have found it conwhere Ho(zl ,z 2) is the transfer function corresponding to venient to classify the shadow characteristic as soft, medium, (27) and is given by or harsh, depending upon the value of yi. For the three cases described in Table I, the corresponding power-spectral densities Sr,(f2) and S & ) together with the i?, H o ( z l) = ,z2 bj,jz:iz;/ t ai,jz;iz;i- , resulting Wiener filter response Ho(S2>,are plotted in Fig. 9 (WED (i, ) E D ' j as a function of the radial frequency variable 5 . Observe the 2 (29) sharp dc rejection in all cases. This is, of course, due to the In choosing Ho(zl ,z 2 ) to provide an approximation to Ho(CL) relatively high concentration of energy at low spatial frequenfor zi = i = 1, 2, wehave restrictedattention to the case cies associated with the illuminationprocess fi(x>,as described previously. The Wiener filter response Ho(52) exhibits the where Ho (zl ,z 2 ) is a second-order section of the form same filter recursing from each of the four corners. If {hi,j} represents the point spread function associated with a single application of the filter specified by (27), then the composite filter possesses point spread function as indicated in Fig. 8. The corresponding system transfer function of the composite filter is then

[1

restriction the with

boo = -2b01
-

bll

2boz.

(31)

This choice ensures zero frequency response at the origin and symmetry of the corresponding point spread function about a

essential characteristics of the linear high-pass filter originally proposed by Stockham [4] on anad hoc basis. It is of some interest that this choice can be justified rigorously under the modeling assumptions madehere. In Fig. 10 we illustrate three-dimensional plots of one quarter-plane of the frequency response of thefilters asso-

FRIES AND MODESTINO: IMAGE ENHANCEMENT

BY STOCHASTIC HOMOMORPHIC FILTERING

633

-20

OD1

004 005 Radial Frequency 0/2t

0302 003

0.06

- 20 0

z -10 O
-20
0.01

002

0.03 0.04 0.05

Rodial Frequency

n/2r

0.06

'

00 .1

0 0 0 0 0.04 a05 .2 .3 Radial Frequency 0 / 2 r

0.06

(a) Fig. 9. Illustration of typical power-spectral densities and the associated Wiener filter response. (a) Softshadow model. (b) Medium shadow model. (c) Harsh shadow model.

(C)

(f) (e) Fig. 10. Frequency responses of desired and actual filters. (a) Desired response for soft shadow. (b) Actual response for soft shadow. (c) Desired response for medium shadow. (d) Actual response for medium shadow. (e) Desired response for harsh shadow. (f) Actual response for harsh shadow.

ciated withthethree cases described in Table I. Thelefthandcolumn shows the desiredresponses while the righthand column illustrates the responses obtained by the digital filters. The frequencies displayed run from 0 at the front to R radians per pixel at the back. Note that in all cases the complete attenuation dc of and rolloff at higher frequenceis has been preserved in the transition from the desired analog filters to the digital approximations.

VII. RESULTS The algorithm presented above has been applied to several real-world images for selected values oftheparameters y n , yi, p,, X p i , and Xi. The processed images shown here illus,.,

tratethe effects of varying theseparametersonthereconstructed images. Parts (a) of eachof thefour sets of images comprising Figs. 11-14 are the original images, while parts (b), (c), and (d) are the corresponding processed versions.. All of the processed images have the salt-and-pepper noise 20 dB below the reflectance, while the reflectance has zero correlation of intensities in different regions ( p , = 0), and an edge density such that on the average there will be 12 transitions on a side of the image (X, = 0.05). The variations among (b), (c), and (d) are due to changes in the illumination intensity y and the illumination i correlation pi. Part (b) of these figures has the edge density of the illumina-

6 34

ILEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-27, NO.

6 , DECEMBER 1979

Fig. Fig. 1 1 . Typical results on the head-and-shoulder image. (a) Original. (b) Soft shadow. (c) Medium shadow. (d) Harsh shadow. shadow.

12. Typical results onanoutdoor (c) Medium shadow.

scene. (a) Original. (b) Soft (d) Harsh shadow.

FRIES AND MODESTINO: IMAGE ENHANCEMENT BY STOCHASTIC HOMOMORPHIC FILTERING

635

Fig. 13. Typical results on abuildingimage.(a)Original. (b) Soft shadow. (c) Medium shadow. (d) Harsh shadow.

Fig. 14. Typical results on an FLIR image. Original. (a) (b) shadow. (c) Medium shadow.(d) Harsh shadow.

Soft

636

IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-27, NO. 6 , DECEMBER 1979

(4

(dl

Fig. 15. Typical results on a tomograph. (a) Original. (b) Soft shadow. (c) Medium shadow. (d) Harsh shadow.

tion process Xi = 0.0125 or, on the average, three transitions on a side of the image, pi = 0.5 (positive correlation of the illumination in adjacentregions)andillumination power 6 dB below reflectance power ( y j = -6 dB), corresponding to a soft shadow. Note how in part (b) local variations are much more evident; shadow and highlight detail have both been emphasized. Specifically, inFig. ll(b) detail in the hair is muchmore evident than in Fig. 1l(a). In Fig. 12(b) the light scaffolding blends in less with the clouds, while the dark slatting on the side of the tower is also more evident. Similarly, in Fig. 13(b) the bright regions in the center have been deemphasized and detail in the building is more apparent. Finally, for the forward looking infrared (FLIR) image Fig. in 14(b), the tank blends in less with its surroundings. This processing brings out detail in both dark andlight areas. Parts (c) (medium shadow) of Figs. 11-14 vary from parts (b) in the increaseof theilluminationpowerby3 dB.The effects noted above have simply become more evident. Part (d) (harsh shadow) deviates from (c) in both illumination power (which is increased by 9 dB) and the correlation of illuminances in different regions which is now -0.9 instead of 0.5. Part (d) of all the figures brings out detail in adjacent light and dark regions better than (c) as is expected from the shadow being modeled as having black-to-white transitions across boundaries between adjacent regions. An application of this technique to a dynamic range reduction problem is illustrated in Fig. 15. In the original

tomograph image little detail is evident in the dark area, while the light areas appearmuddy.In Fig. 15(b) some detail is evident in the dark region, while local shading is more apparent in light regions. When harsher shadows are assumed, as in 15(c) and 15(d),local detail is emphasized. In allcases the processing hasincreasedthe visibilityof detail in such a way that less scrutiny need be applied when observing the processed images than is required when observing the originals. There are, however, differences in the characteristics of the processing that occurs. Increasing the shadow power and making the shadow more pronounced by having negative correlation (pi= -0.9) instead of positive correlation (pi = 0.5) has made local detail more visible by suppressing variations in intensity that occurover large regions.
AND CONCLUSIONS VIII. SUMMARY

A method for determining thelinear filtering operation to be performed as part of homomorphic processing based on Wiener filtering conceptshasbeendescribed.The resulting linear filter is determined by the parameters of the assumed stochastic models of the signal and degradation. These parameters are associatedwith salient characteristics of the signal and degradation. Although no claims can be made about the optimality of a particular filter for any given real-world image, it has been shown that the parameters in the models provide a useful meansfor specifyingthe desired effects of the processing. The implementation of the linear operation has taken the

IEEE TRANSACTIONS

ON ACOUSTICS, SPEECH,

AND SIGNAL PROCESSING, ASSP-27, VOL. NO.

6, DECEMBER 1979 637

formtwo-dimensional of recursive IIR digital filter whose order has been restricted to result in computational efficiency and whose coefficients have been chosen to result in a good approximation to the Wiener filter. The resulting nonlinear filtering operationhasbeenapplied to a number of real-world images to illustrate the relationship between the parameters of the assumed stochastic models and the effects of the resulting processing.

ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers who helped to improve earlier version of this paper. an

REFERENCES
[ 1 W. D. Zoethout, PhysiologicalOptics. Chicago, IL:The Profes1 sional Press, 1939, p. 157. New York: Wiley, [2] P. Kowaliski, AppliedPhotographicTheory. 1972. [3]J. E. Hochberg, Perception. Englewood Cliffs, NJ: PrenticeHall, 1964. [4] T. G. Stockham, Image processing in thecontext of a visual mode1,Proc. IEEE, vol. 60, pp. 828-842, July 1972. [SI H. L. Van Trees, Detection, Estimation and Modulation Theory; Vol. 1. New York: Wiley, 1968. [6] R. W. Fries, Theory and application of a classof two-dimensional random fields, Ph.D. dissertation, Rensselaer Polytech. Inst., Troy, NY, in preparation. [7] A. V. Oppenheim, R. W. Schafer, and T. G. Stockham, Jr., Nonlinear filtering of multipliedand convolved signals, pfoc. ZEEE,vol. 56, pp. 1264-1291, Aug. 1968.

E. Wong, Homogeneous Gauss-Markov randomfields, Ann. Math. Stat., vol. 40, pp. 1625-1634,1969. [91 E. Wong, Two-dimensional random fields and the representation ofimages, SIAM J. Appl.Math., vol. 16, pp. 756-710, 1968. E.Wong, Stochastic Processes in InformationandDynamical [lo] Systems. New York: McGraw-Hill, 1971, ch. 7. [111 S . Bochner, Lectures on Fourier Integrals, Annals o f Mathematical Studies, Vol. 42. Princeton, NJ: Princeton Univ.Press, 1959, pp. 235-238. Optical systems, singularity functions, complex [I21 A. Papoulis, Hankel transforms, J. Opt. Soc. Amer., vol. 57, pp.207-213, 1967. with i n ~ 3 A. Papoulis, Systems and Transforms Applications 1 Optics. New York: McGraw-Hill, 1968. [ 141 J. W. Modestino and R. W. Fries, A generalization of the random telegraph wave, submitted to ZEEE Trans. Inform. Theory. -, Edge detectionin noisyimagesusingrecursive digital filtering, Computer Graphics and Image Processing, vol. 6, pp. 409-433, Oct. 1977. [ 161, J. W. Modestino, R. W. Fries, and D.G. Daut, A generalization of the two-dimensional random checkerboard process, to appear in J. Opt. SOC.Amer. ~ 7 J. W. Modestino and R. W. Fries, Construction and properties 1 of a useful two-dimensional random field, submitted to IEEE Trans. Inform. Theory. [18] L. R. Rabiner and B. Gold, Theory and Applications o f Digital Cliffs, Prentice-Hall, NJ: 1975, Signal Processing. Englewood p. 5 1. ~ 1 9 1C. Shipman, Understanding Photography. Tuscon, AZ: H. P . Books, 1974. [ 201 J. Mannos and D. L. Sakrison, The effects of a visual fidelity IEEE Trans. Inform. criterion the on encoding of images, Theory, vol. IT-20, pp. 525-536,1974. [ 211 H. C. Andrews and B. R. Hunt, Digital Image Restoration. Englewood Cliffs, NJ: Prentice-Hall, 1977.

PI

A New Approach for the Design of Digital Interpolating Filters

GEERD OETKEN

Aktmct-The paper presents a new class of interpolators characterized by the property that themean power of their error sequence, a as function of frequency, approximates zero inthe Chebyshev sense. The design method, which is based on some observation and newly found general properties,will be describedin some detail. An example shows thespecialproperties of theseinterpolatorsin comparison with former results. Measurements done with a practical implementation are presentedas well. A design chart for these filters is provided.

based on an error function, the definition of which repeated is here for convenience (see Fig. 1). The given sequence { ~ ( k ) } , be interpolated with a rater, to is considered as being a sampled version of the unknown, but bandlimited, sequence uo(k). For a convenient notation, we disregard causality for the time being and write the transfer function of theFIR system to bedesigned as
N

H(2) =

h(2)z-k
k=-N

INTRODUCTION HE design of digital interpolators has been consideredin quite a few recent papers (e.g., [ I ] - [ 6 ] ) . In this paper we refer especially to [ 2 ] and [ 3 ] ,where the investigation is
I.

Manuscript received August27,1975; revised October 15,1975, April 26,1976, and April 16,1979. The author is with the Institut fur Nachrichtentechnik, Universitat Erlangen-Niirnberg, West Germany.

where n = 2 N = 2Lr is the degree of the system, and 2 L the number of input samples the interpolation is based on. Obviously the output sequence is y ( k ) = h(k) * u(k). Now we apply a sinusoidal input sequence as a test signal, written as uo(k) = Uo(Q) ejkn with Vo(S2)= 0, Q 2 Qg = m / r , 0 < a < 1 , Using the error sequence A y ( k , Q) = y ( k )- u,(k), an error function was defined in [ 2 ] and [ 3 ] as

0096-3518/79/1200-0637$00.75 0 1979 IEEE