Image restoration, enhancement and target

location with local adaptive linear lters

Leonid Yaroslavsky
Department of Interdisciplinary Studies, Faculty of Engineering, Tel Aviv
University, Tel Aviv 69978, Israel

Abstract. Local adaptive lters for image restoration (denoising and deblurring),
enhancement and target location local are described. The lters work in the do-
main of an orthogonal transform ( DFT, DCT or other transforms) in a moving
window and nonlinearly modify the transform coecients to obtain an estimate of
the central pixel of the window. A framework for the lter design for multi compo-
nent images is presented and experimental results in denoising, enhancement and
deblurring monochrome and color images and in target location are provided. Also
discussed is lter implementation for the window size of 3  3 pixels in a form
of 5 channel convolution with masks implementing local mean and 4 directional
Laplacians: horizontal, vertical and two (45 ) diagonal.
Keywords: Image restoration, Image enhancement, Target location, Adaptive
lters, Transforms

1 INTRODUCTION
Local adaptive linear lters for image restoration and target location that are
the subject of the present paper implement the idea of space (time)-frequency
signal representation that dates back to 40-th ( [1]) though it has not became
practical (at least in image processing) until 80-th ([2]) when the computer
technology provided computers with an appropriate speed and memory size.
The lters work in a moving window in the domain of an orthogonal transform
and, in each position of the window, nonlinearly modify the signal transform
coecients to obtain an estimate of the central pixel of the window.
Image processing in transform domain rather than in signal domain sug-
gests certain advantages in terms of the convenience of incorporating a priory
knowledge on images into the design of processing algorithms and in terms of
computational expenses. The transfer from signal domain into the transform
domain is especially promising if it is applied locally rather than globally.
In this paper, we overview basic ideas for the lter design and implemen-
tation, justify using for this purpose Discrete Cosine Transform and present
experimental results in denoising and deblurring monochrome and color im-
ages and in target location. We then discuss using other than DCT orthogo-
nal transforms such as Walsh-Hadamard and Haar transforms and show also
that, for the window size 3  3 pixels, ltering in DCT transform domain
is equivalent to 5 channel convolutions with masks implementing local mean
and 4 directional Laplacians: vertical, horizontal and 45 diagonal ones.
2 Leonid Yaroslavsky
2 MULTI COMPONENT LOCAL ADAPTIVE
FILTERS
2.1 Optimal local adaptive lters
For the design of an optimal lter one should formulate an optimization
criterion. The design of local adaptive lters is based upon local criteria that
evaluate image processing quality for each particular image sample as an
average, over a certain neighborhood of the pixel, of a certain loss function
that measures deviation of the estimation of each pixel in the neighborhood
from its true value([2,4{7]). In what follows we assume that images to be
processed are multi component and represented by their components such as
R, G, andn (Bc) components
o n (cof) color
o images.
Let an ;n and ban ;n are samples of ideal image components and
1 2 1 2

their estimations, respectively,fn1; n2 g are pixel coordinate indices, C is the

number of image components, (k1 ; k2 ) are running coordinates in the im-
age plane. Let also LOC(a(kc);k ; n1 ; n2 ) is a locality function that denes the
pixel's ak ;k neighborhood and is nonzero only for those pixels with coordi-
1 2
1 2

nates fn1; n2 g that belong to the neighborhood and let LOSS(a(nc);n ; ba(nc);n ) 1 2 1 2
is a loss function that describes residual degradation of image samples after
the processing. Then, for multi component images, the local criterion can be
expressed in the following form:
X
C X
AVLOSS (k ; k ) = AVimsys (AVobj (
1 2 LOC(akc ;k ; n ; n )
( )
1 2 1 2
c=1 fn1 ;n2 g
 LOSS(anc ;n ; banc ;n )))
( )
1 2
( )
1 2
(1)
where AVimsys (:) denotes operation of averaging over realizations of imaging
system sensor's noise and AVobj (:) denotes averaging over unknown param-
eters of the image (such as imaged objects' position, orientation, size) that
arenot known before the processing. The design of optimal local adaptive
n (c) iso aimed at nding a mapping
lters n (ofc) theo set of observed image samples
bn ;n to the set of estimations ban ;n that provides minimum to the
1 2 1 2

criterion for each pixel of the image. For multi component images, if the
neighborhood is dened as a spatial one
 jk n j  M ; jk n2 j  M2 ; 8 c (2)
LOC(n ; n ; c; akc :k ) = 10;; ifotherwise
1 2
( )
1 2 :
1 1 1 2

and the loss function is a quadratic one:

 
LOSS banc ;n ; anc ;n = jbanc ;n
( )
1 2
( )
1 2
( )
1 2
a(nc1);n2 j2 ; (3)
 Local adaptive ltering 3
the local criterion requires minimization of the functional:
AVLOSS(k ; k ) 8 1 2
9
>
> >
>
>
<X C X >
=
= AVimsys AVobj > jbanc ;n ( ) c
an ;n j
( )
>
2
(4)
> >
1 2 1 2
c
: jk n j  M
> >
;
=1
1 1 1
jk 2 n2 j  M2
over all mappings of the observed signal to its estimation. To make the
design constructive, one should parameterize the observation-to-estimation
mapping. As an option, one may assume that the estimation of each pixel
is obtained as a weighted sum of the observed pixels in its neighborhood. In
this way we arrive at local adaptive linear lters.
For local adaptive lters, lter coecients have to be determined in each
particular position of the running window from the observed image samples
within the window on the base of a priori information regarding images un-
der processing. While formulation of this information in the image domain
is apparently problematic, it is much simplied in the domain of certain or-
thogonal transforms, such as DFT and DCT, because power spectra of image
fragments in such transforms exhibit much more regular behavior than that
of pixels themselves. For instance, it is well known, that image DFT and
DCT spectra usually decay more or less rapidly to high frequencies.
Therefore we will assume in what follows ltering in a transform domain
and will design lters that operate, in each position (k1 ; k2 ) of the window,
in the following 3 steps: n o
1. Computing spectral coecients r(k;r;k;) = Tb(k ;k ) of the observed1
1
2
2 1 2

image fragment b(k ;k ) within the window over the chosen orthogonal trans-
1 2

form T.
n (k ;k )o of the obtained spectral coecients by the lter coe-
2. Multiplication
cients r ;r ; : 1
1
2
2

b(rk11;r;k2;
2)
= r(k1 ;r1 ;k2 ;2 ) r(k1 ;r1 ;k2 ;2 ) . (5)
n o transformation T 1 of the output signal spectral coecients
3. Inverse
br ;r ; to obtain an estimated value of the central pixel of the window.
k
( 1
1
;k )
2
2

Here, subscripts (r1 ; r2; ) are corresponding indices in the transform domain.
With this approach, the synthesis of local adaptive ltersn is(kreduced
;k o to the
determination of (2M1 + 1) (2M2 + 1) C lter coecients r ;r ; . For the
)
1
1
2
2

optimal lter design in the domain of an orthogonal transform, one can, by

virtue of the Parceval's relation, reformulate the criterion of Eq.(4) in terms
of signal spectra :
4 Leonid Yaroslavsky

AVLOSS (k ; k ) ( 1 2
)
XX
C
= AVimsys AVobj jrk1 ;r;k2 ;2 ) r(k1 ;r1 ;k2 ;2 )
( 1
rk1 ;r;k2 ;
( 1
j
2) 2
(6)
r1 ;r2 =1
n o
By minimizing AVLOSS (k1 ; k2 ) with respect to r(k;r;k;) one can nd that 1
1
2
2

the optimal values of the coecients of the lter that minimize the ltration
error as dened by Eq.(4) may be found from the following equation:
n  ?o
AVimsys AVobj rk;r;k; rk;r;k; ( 1 2) ( 1 2)
k ;k
r ;r ; =
( 1 2) n o 1 2 1 2
(7)
1 2
AVimsys AVobj jrk;r;k; j ( 1
1 2
2) 2

with ? denoting complex conjugate. The design of the local adaptive lter of
Eq.(7) is therefore reduced to an estimation of local power spectrum of the
input image fragment and its mutual local spectrum with the 'ideal' image.

2.2 Local adaptive lters for image restoration

Assume that image distortions can be modeled by the equation:
b = La + n; (8)
where L is a linear operator of the imaging system and n is a random zero
mean signal independent random vector that models imaging system sensor's
noise. Assume also that the imaging system operator L is such that the
distorted image can be described in the domain of the chosen orthogonal
transform by the following relationship:
r(k1 ;r1 ;k2 ;2 ) = (rk11;r;k2 ;
2 ) (k1 ;k2 )
r1 ;r2 ; + r(k1 ;r1 ;k2 ;2 ) (9)
n o
where (rk;r;k;) are running representation coecients of the linear operator
1 2
1 2
n o
L in the domain of the orthogonal transform and r(k;r;k;) are zero mean 1
1
2
2

spectral coecients of the realization of the noise interference. Then one can
obtain from Eq.(7) that optimal restoration lter coecients are dened as:

r(k1 1;r;k2 ;
2)
= jrk;r;k; j jrk;r;k; j
( 1
1 2
2) 2 ( 1
1 2
2) 2

(10)
rk;r;k; AVimsys AVobj jrk;r;k; j
( 1
1 2
2) ( 1
1 2
2) 2

Filter of Eq.(10) can be regarded as an "empirical" Wiener lter that assumes

estimation, in each position of the window, of the parameters involved in its
design locally from the observed distorted image fragments.
 Local adaptive ltering 5
2.3 Local adaptive lters for image enhancement
Image enhancement is a processing aimed at assisting visual image analy-
sis. A reasonable basis for the design of linear lter parameters for image
enhancement is an assumption that the ltering should "restore" a "useful"
signal, that is image details to be enhanced or extracted for the end user's
convenience, against a "noise" background that obscures interpretation of the
details. Following the above approach, one can, by analogy with the deriva-
tion of Eq.(10), design a lter that provides minimum to the average squared
modulus of the dierence between this useful signal and the ltered signal:
 
AVobj jrk;r;k; j ( 1 2) 2
k ;k
r ;r ; =
( 1 2)   1 2
(11)
AVimsys AVobj jrk;r;k; j
1 2 ( 1
1 2
2) 2

where AVobj denotes averaging over such variations of the useful object(s)
as object position, size, shape, etc.
Another useful criterion for image enhancement is the criterion of signal
spectrum restoration ([12]) that requires restoration of power spectrum of
the object signal. For this criterion, the following obvious lter results:
0  k ;k  1 =
AV
1 2
obj jr ;r ; j ( 1 2) 2

rk;r;k; = @  k ;k  A
1 2
( 1 2)
(12)
1 2
jr ;r ; j
( 1
1 2
2) 2

2.4 Local adaptive lters for target location

In target location, the target position is found by locating the signal highest
maximum at the output of a linear lter ([7,8]). Therefore the lter design has
to be aimed at achieving maximal ratio of the lter output signal at the target
actual position to the highest signal maximum in the eld of search that is
over the background component of the image fragment within the window.
To all practical purposes, this requirement is equivalent to the requirement of
maximizing the ratio of the lter response to the target signal to the standard
deviation of the signal over the eld of search ([8,10]).
In this way one can arrive at the following optimal lter for target location:
 ?
(rk11;r;k2 ;
2)

r(k1 1;r;k2 ;
2)
=  ; (13)
AVimsys AVobj jrk;r;k; j ( 1
1 2
2) 2

 ?
where (rk;r;k;) is complex conjugate to the spectrum of the target signal,
1
1
2
2

jr(k;r;k;) j2 is power spectrum of the image background component and AVobj

1
1
2
2

in this particular formula denotes averaging over unknown position of the

target.
6 Leonid Yaroslavsky
3 SELECTION OF THE TRANSFORM
The selection of orthogonal transforms for the implementation of the lters
is governed by the convenience of formulating a priori knowledge regarding
image spectra in the chosen base, by the accuracy of spectrum estimation
from the observed data that is required for the lter design, and by the
computational complexity of the lter implementation.
The transforms that seem to be the most appripriate in these respects are
the Discrete Fourier and Discrete Cosine Transforms. DCT is advantageous
to DFT in terms of the accuracy of spectral estimation due to the fact that
DCT, being DFT of signals evenly extended outside their borders, substan-
tially eliminates boundary eects that are characteristic for the DFT proper
because of its periodicity. It is illustrated in Figs. 1 and 2, where one can
easily see that local DCT spectra converge to zero much faster than those of
DFT. An important computational advantage of DFT and DCT is that both,

Fig. 1. Comparison of local DFT and DCT spectra shown in coordinates of spectral
index (vertical) versus signal sample index (horizontal) for a test chirp signal.

DFT and DCT in a running window can be computed recursively ([2{5,7]).

 Local adaptive ltering 7

Fig. 2. Comparison of local DFT and DCT spectra for an electro cardiogram test
signal.

Therefore, the computational complexity of ltering in the base of DFT and

DCT in a running window is of the order of O(Size of the window) operations
which is the theoretical minimum for a general space variant ltering. In the
case of ltering in DCT domain with odd size of the window, in the computa-
tion of the inverse transform for the central pixel of the window only spectral
coecients with even indices are involved. Therefore, only these coecients
have to be computed and modied in the ltering in the DCT domain in a
moving window, and the computational complexity of such a ltering can be
asymptotically estimated as O (WindowSize=4).
As for the component-wise transform, if we assume that it is separable
from the transforms over spatial coordinates, it will not aect the computa-
tional complexity. We will not specify it here and leave the question of the
selection of it open. An illustrative ow diagram of local adaptive lering in
DCT domain is shown in Fig. 3.
8 Leonid Yaroslavsky

Fig. 3. Flow diagram of local adaptive ltering in transform domain

4 FILTER IMPLEMENTATION: LOCAL ADAPTIVE

FILTERS WITH NONLINEAR PROCESSING IN
TRANSFORM DOMAIN
The main issue in the design and implementation of local adaptive lters in
the transform domain is the estimation, in each position of the window, of
averaged spectra of the observed image fragments, of the desired signal and
of the signal fraction regarded as an additive signal independent noise or as
a background. In this estimation, a priori knowledge regarding image and
 shouldbe utilized. In image restoration, noise power spectrum
noise spectra
AVimsys jr(k;r;k;)j2 , can either be known from the imaging system design
1
1
2
2

or can be estimated from observed noisy images ([7]). The estimation

 of the
(k ;k ) 2
observed image fragment spectrum AVimsys AVobj jr ;r ; j may be car- 1
1
2
2

ried out by one or another smoothing the observed spectrum of the fragment
being processed. The estimation of the spectrum of "ideal" image fragments
in the case of additive signal-independent noise can then be carried out using
the relationship:
jrk;r;k; j jrk;r;k; j '
( 1 2) 2 ( 1 2) 2
1
n 2 1 2
o
max 0; AVimgsys AVobj jrk;r;k; j ( 1
1 2
2) 2
AVimgsys jrk;r;k; j
( 1
1 2
2) 2
; (14)
 Local adaptive ltering 9
As a zero order approximation, one can use the observed power spectrum
jr(k;r;k;) j2 as an estimate of the averaged one AVimgsys AVobj jr(k;r;k;) j2 . In
1
1
2
2
1
1
2
2

this way we arrive at the following implementation of lters for image de-
noising and deblurring:
8  k ;k 
< j  r ;r ; j jr ;r ; j
k ;k
if rk;r;k; 6= 0 ;
( 1 2) 2 ( 1 2) 2

rk;r;k; = max 0; rk ;r;k; jrk ;r;k; j

1 2 1 2 ( 1 2)
( 1
1
: 2
2) ( 1 2)
1 2
( 1 2) 2
1 2
1 2
(15)
0; otherwise
Similar implementations can be obtained for lters of Eqs.(11-13) for image
enhancement as well. The two following modications of this lter can also
be suggested:
a "rejective" lter
 (k1 ;k2 ) 2
r1 ;r2 ; = 1=r1 ;r2 ; ; if jr1 ;r2 ; j  thr and r1 ;r2 ; =
k ;k
( 1 2) 6 0; (16)
0; otherwise
where the value of thr is associated with the variance of additive noise, and
a "fractional spectrum" lter
 (k1;k2 ) P 1 (k1 ;k2) 2
r1 ;r2 ; = Gjr1 ;r2 ; j ; if jr1 ;r2 ; j  thr
k ;k
( 1 2)
(17)
0; otherwise
with P as a spectrum enhancement parameter and G as an energy normal-
ization parameter. When P = 1, the lter is equivalent to that of Eq. 16.
The selection P  1 results in signal energy redistribution in favor of weaker
(most frequently, higher frequency) components of local spectra. This mod-
ication is useful for image blind deblurring and image enhancement. Note
that the idea of shrinkage of transform coecients that are lower that a cer-
tain threshold recently reappeared and obtained a popularity in the form of
wavelet shrinkage ([13]).
Note furthermore that if the threshold in the rejecting lter is made an
appropriate function of the signal dc component (for instance it is made
proportional to it or to its square root), the lter can be used for suppression
signal dependent noise such as for instance speckle noise.
The design of lters for target location requires, as it follows from Eq.(13),
estimation of power spectrum of image fragment background component
jr(k;r;k;) j2 averaged over the sensors's noise and over unknown coordinates
1
1
2
2

of the target. As it was shown in Ref. [10] a reasonable approximation to this

estimation is
jr(k;r;k;) j2 = jr(k;r;k;) j2 + j(rk;r;k;) j2
1
1
2
2
1
1
2
2
1
1
(18)
2
2

such that the optimal local adaptive lter can be implemented as

 ?
(rk11;r;k2 ;
2)

2) 
r(k1 ;r1 ;k2 ; = (19)
jrk;r;k; j + jrk;r;k; j
( 1
1 2
2) 2 ( 1
1 2
2) 2
10 Leonid Yaroslavsky

Fig. 4. Upper row: initial piece-wise constant and noisy test images (standard devi-
ation of noise is 20 in the range 0-255). Bottom row: image ltered by the rejective
lter of Eq. (16) with window size is 7  7 pixels , r1 ;r2 ; = 1, thr = 20 and a
map of local lter transparency. Pixel gray level on this map is proportional to the
number of signal spectral components that are preserved by the lter.

Fig. 5. Graphs of the video signal of images in Fig. 4

 Local adaptive ltering 11
Image denoising, deblurring and enhancement capability of the described
local adaptive lters is illustrated on gures 4 - 9. Fig. 4 and graphs on
Fig. 5 illustrate ltering for suppressing additive Gaussian noise in a single
component test piece-wise constant image. From the map of the local lter
transparency (Fig. 4, right bottom corner) and from graphs in Fig. 5 one can
see that the lter is automatically context sensitive: within uniform areas it
acts almost as a local mean lter which passes through practically only local
dc component while in the vicinity of edges it is practically transparent to all
local frequency components of the image and therefore preserves the edges
(at the expense of not suppressing noise). Fig. 6 illustrates denoising a color
image. Fig. 7 represents an example of blind restoration of a blurred color
image. Fig. 8 shows an example of image enhancement. Fig. 9 illustrates the
use of local adaptive lter for target location in stereoscopic images.

Fig. 6. Denoising additive Gaussian noise in a color image. Standard deviation of

noise in each of 3 components was 20 within signal dynamic range 0-255. Left half of
the image represents noisy image, right half - ltered one. Filtering was performed
over right half of the image by the lter of Eq. (16) in the window 7  7 pixels with
(rk11;r;k2 2;) = 1; and thrr = thrg = thrb = 20
12 Leonid Yaroslavsky

Fig. 7. Blind image restoration: initial (left) and restored (right) images. Filtering
was performed with the lter of Eq. (17) in the window 7  7 pixels with P =
0:75; G = 2 and thrr = 50; thrg = 30; and thrb = 60:

Fig. 8. Single component image enhancement by lter of Eq. (17) in the window
15  15 pixels with P = 0:5, thr = 15 and G = 3: initial (left) and ltered (right)
images
 Local adaptive ltering 13
The result of localization (marked with a cross) Target (highlighted)

Correlators output crossection (column) Correlators output crossection (row)

40 40

20
20

0
0
20
20

50 100 150 200 250 20 40 60 80 100 120

Fig. 9. Localization of a small (8  8 pixels, highlighted) fragment of the right of

two stereoscopic images on the left image by the local adaptive lter of Eq.(19).
Graphs illustrate dierence between lter response to the target (sharp peak) and
that to the rest of the image. Running window size is 32  32 pixels; image size is
128  256 pixels. The same lter applied globally rather than locally fails to locate
the fragment.

5 THE USE OF OTHER TRANSFORMS

Although DCT and DFT are advantageous to other transforms in terms
of the convenience of incorporating a priori knowledge on images and their
distortions into the lter design and in terms of the computational complexity
of the lter implementation because they allow recursive spectral analysis in
the moving window, some other transforms can also be used for local adaptive
ltering in transform domain. The closest candidates are transforms with
binary basis functions such as Walsh-Hadamard and Haar Transforms that
are computation wise the most simple ones. For moderate window sizes, basis
functions of DCT and Walsh-Hadamard transforms dier from each other not
too much as one can see from Fig. 10. Fig. 11 illustrates denoising of a 1-
D signal of an electro cardiogram by local adaptive rejective ltering in the
Haar Transform domain.
14 Leonid Yaroslavsky

Fig. 10. Arrais of 8  8 basis functions of (from left to right) 2-D DCT, Walsh-
Hadamard and Haar Transforms of the order 8.

Fig. 11. Filtering noisy electrocardiogram in the domain of Haar Transform.

 Local adaptive ltering 15
6 MODIFICATION OF THE DENOISING
PROCEDURE: THRESHOLDING THE
DIRECTIONAL LAPLACIANS.
Described local adaptive lters are context sensitive. They successfully sup-
press noise in areas where video signal is changing slowly while in the vicinity
of edges and other heterogeneous areas they tend to keep signal unchanged.
Selection of the lter window size depends on what type of image details
should be preserved and what can be neglected.
In the restoration of high resolution images when objects of the size of
about one pixel should be restored the lter window size should be minimal
in order to preserve those tiny details. Minimal window which preserves sym-
metry with respect to its central pixel is 3  3 pixel window. Let us analyze
basis functions of local DCT spectral analysis in such a window. These 9
basis functions are:
2 DCT 00 DCT 10 DCT 20 3
4 DCT 01 DCT 11 DCT 21 5 =
DCT 02 DCT 12 DCT 22
2 2 1 1 1 3 2 p3 0 p3 3 21 3 3
2 1
66 4 1 1 1 5 4 pp3 0 pp3 5 41 2 1 5 77
66 1p 3 77
66 2 p3 p13 1 p13 3 2 3 3 00 33 3 2 p 1
3 2 3
p2
3 777
66 4
66 p3 p3 p3 5 4 0 3 00 03 5 4 0p p0 p0 5 77
33 77
(20)
66 2 1 1 1 3 2 p3 0 p3 3 23 12 3 2 1 77
64 4 2 2 2 5 4 2p3 0 2p3 5 4 24 25 5
1 1 1
p p 1 2 1
3 0 3
It follows from the denition of 2-D DCT, that from them, only function
DCT 00; DCT 20; DCT 02 and DCT 22 are involved in the computation of
the window central pixel value. These functions being applied in a running
window generate image local mean (function DCT 00) and image horizon-
tal (function DCT 02), vertical (function DCT 20) and isotropic (function
DCT 22) Laplacians. The latter can be decomposed into a sum of Laplacians
taken at 45 :
DCT 22 = (DCT 220 + DCT 221) (21)
where
2 2 1 1
3 1 1 2
2 3
DCT 220 = 4 1 2 1 5 and DCT 221 = 4 1 2 1 5 (22)
1 1 2 2 1 1
16 Leonid Yaroslavsky
This casts a new light at the lter denoising capability and leads to dierent
modications of the lter. One of the possible modications is a direct im-
age 5-channel convolution with window functions DCT 00; DCT 20; DCT 02,
DCT 220 and DCT 221; thresholding the convolution results and combining
them into the output image.
Second possible modication is a "soft thresholding" of the convolution
results:
 INP UT; if jINP UT j > thr
OUT P UT = thr
sign (INP UT )
j INPUT jP
; otherwise (23)
thr
where thr is the same threshold parameter as that in Eq. 16 and P is a
parameter that denes the degree of approximation of the "hard threshold-
ing" of Eq. 16 (the "hard thresholding is achieved when P ! 1): The "soft
thresholding" may help to reduce artifacts of denoising caused by the loss of
low contrast image details. Fig. 12 illustrates image deblurring and denois-
ing with the use of the above 5-channel convolution with window functions
DCT 00; DCT 20; DCT 02, DCT 220 and DCT 221 and the subsequent "soft"
thresholding.

Fig. 12. Correction of image sensor scanning aperture by 5-channel convolution

with directional Laplacians. Left to right: initial image and aperture corrected im-
ages withot anf with noise suppression.

Further natural modications are shaping the Laplacian's masks DCT 00,
DCT 20, DCT 02, DCT 220, and DCT 221 by an appropriate window function
and a directional smoothing the convolution results. The shaping may help to
reduce artifacts associated with the rectangle sampling raster. The directional
smoothing may provide better denoising capability than the simple element-
wise thresholding.
 Local adaptive ltering 17
7 CONCLUSION
We have described local adaptive lters for image restoration, enhancement
and target location that work in a moving window in the domain of the
DCT and other transforms, outlined the lter design and implementation
and demonstrated their applications for processing 1-dimensional signals,
monochrome and color images. We also have shown that ltering in the DCT
domain by the local adaptive ltering in the window 3x3 is equivalent to
image convolutions with uniform aperture and horizontal, vertical and two
diagonal (45) 3  3 Laplacians.

References
1. Gabor D. (1946), Theory of Communication, J. Inst. Elec. Eng.,93: 429-457
2. Yaroslavsky L. (1987) Digital signal processing in optics and holography, Radio
i Svyaz', Moscow (in Russian)
3. Vitkus R. Yu., Yaroslavsky L.P.(1988) Adaptive Linear Filters for Image Pro-
cessing, In: V.I. Siforov, L.P. Yaroslavsky (Eds.) Adaptive Methods for Image
Processing, Nauka Publ.,Moscow: 6-35. (In Russian)
4. Yaroslavsky L. P. (1991) Linear and rank adaptive lters for image processing.
In L. Dimitrov and E. Wenger, (Eds) Digital Image Processing and Computer
Graphics. Theory and Applications, Oldenburg, Wien{Mnchen: R., pp. 374,
1991
5. L.P. Yaroslavsky (1996) Local Adaptive Filters for Image Restoration and En-
hancement. In: M.-O. Berger, R. Deriche, I. Herlin, J. Jare, J.-M. Morel (Eds.)
12-th Int. Conf. on Analysis and Optimization of Systems. Images, Wavelets
and PDE's, Paris, June 26-28, 1996, Springer Verlag, Lecture Notes in Control
and Information Sciences, 219, 31-39.
6. Yaroslavsky L. (1996) Local Adaptive Image restoration and Enhancement with
the Use of DFT and DCT in a Running Window. In: Unser A., Aldroubi A.,
Laine A.F. (Eds.) Wavelet Applications in Signal and Image Processing IV,
SPIE vol. 2825, 2-11.
7. Yaroslavsky L. and Eden M. (1996), Fundamentals of Digital Optics,
Birkhauser, Boston
8. Yaroslavsky L.P. (1993), The Theory of Optimal Methods for Localization of
Objects in Pictures, In: E. Wolf (Ed.) Progress in Optics, v.XXXII, Elsevier
Science Publishers, Amsterdam
9. Yaroslavsky L. (1997) Nonlinear adaptive image processing in transform do-
main,In: 1977 IEEE Workshop on Nonlinear Signal and Image Processing, Sept.
7-11, 1997, Grand Hotel, Mckinac Island, Michigan, USA
10. Yaroslavsky L., Marom E. (1997), Nonlinearity optimization in nonlinear joint
transform correlators, Applied Optics, 36, July: 1-7.
11. Vitkus R. Y. and Yaroslavsky L.P. (1987) Recursive Algorithms for Local Adap-
tive Linear Filtration. In: L.P. Yaroslavskii, A. Rosenfeld and W. Wilhelmi
(Eds.) Mathematical Research, Academy Verlag, Berlin, 34-39
12. Pratt W. K. (1978), Digital Image Processing, Wiley, N. Y.
13. Donoho D.L. and Johnstone I.M (1994) Ideal spatial adaptation by wavelet
shrinkage, Biometrica, 81(3): 425-455