JOURNAL OF THE OPTICAL SOCIETY OF AMERICA

Bayesian-Based Iterative Method of Image Restoration*

WILLIAm HADLEY RICHARDSON
Visibility Laboratory, University of California, San Diego, San Diego, California 92152
(Received 15 September 1970)
An iterative method of restoring degraded images was developed by treating images, point spread func-
tions, and degraded images as probability-frequency functions and by applying Bayes's theorem. The
method functions effectively in the presence of noise and is adaptable to computer operation.
INDEX HEADINGS: Spread function; Image restoration; Deconvolution.
This paper reports the results of applying probability
methods to restoration of noisy degraded images.
Fourier-transform methods of image restorations
2
have
been successful when noise content in the degraded
image is moderate or small. At increased noise levels,
however, Fourier methods have failed to produce
recognizable images.
ASSUMPTIONS
It was assumed that the degraded image H was of
the form H= W*S, where W is the original image, S is
the point spread function, and * denotes the operation
of convolution. It was also assumed that W, S, and H
are discrete probability-frequency functions, not neces-
sarily normalized. That is, the numerical value of a
point of W, S, or H is considered as a measure of the
frequency of the occurrence of an event at that point.
S is usually in normalized form. Units of energy (which
may be considered unique events) originating at a
point in W are distributed at points in H according to
the frequencies indicated by S. H then represents the
resulting sums of the effects of the units of energy
originating at all points of W. In what follows, each of
the three letters has two uses when subscripted. For
example, Wi indicates either the ith location in the array
W or the value associated with the ith location. The un-
subscripted letter refers to the entire array or the value
associated with the array as in W = E i Wi. The double-
subscripted Wi j in two dimensions is interpreted
similarly to Wi in one dimension. In the approximation
formulas, a subscript r appears, which is the number of
the iteration.
DISCUSSION
Given the degraded image H, the point spread func-
tion S, and the requirement to find the original image
W, Bayes's theorem comes readily to mind. In the nota-
tion of this problem the usual form of the Bayes's
theorem' may be stated as the conditional probability
of an event at Wi given an event at Hk,
P(HI I Wi)P(lwi)
P(~i I HA;~ P(~ 'V)
)
= 1 I) 1, (1)
AdP(Hk I Wj)P(Wj);
j={1,J}, k={1,K),
where Hk is for the moment an arbitrary cell of H.
Considering all the Hk and their dependence on all
Ws in accordance with S, we can say
P(Wi))=7 P(WiTkV)= Z P(WiIHk)P(Hk), (2)
k k
since P (W Hk) = P (wiHk)/1P (Hk). Substituting Eq.
(1) in Eq. (2) gives
P(WVO = F P(HkI Wi)P(Wi)P(Hs)
k E P(Hk |W j)P(W j)
In the right side of this equation, the term P(Wi) is
also the desired solution. But in many applications of
Bayes's theorem, when this P(Wi) term is not known,
an accepted practice
4
'
5
is to make the best of a bad
situation and use an estimate of the P(Wi) function to
obtain, from Eq. (1), an approximation of P(WiIHO).
When this practice is applied here, Eq. (3) becomes
P(Hk I Wi)P(Hk)
k E P(Hk IWj)Pr(Wj)
(4)
r={0, 1, } .
This results in an iterative procedure where the initial
Ps(Wi) is estimated. An estimation often used is
Bayes's postulate (also known as the equidistribution of
ignorance), which assumes a uniform distribution so
that Po(Wi) = 1/I or WiT-= W/I.
Equation (4) can be reduced to a more easily
workable form by P(Wi)=WiW/ and P(Hk)=Hk/H
= H/IW, since the restoration is a conservative process
and W = H, and also P(Hk Wi)=P(Si,k)=Si,k/S,
S=, Sj, j={1,J}.
i
Then Eq. (4) becomes
(Si,k/S) - (HA/TW)
k E (Sjk/S) -(Wjr/W)
or
Wi,,1=WiT',Si,kHk
TV, ,.+i= Wi.,. E
k E SjkWj.,
i
(5)
As this stands, investigation shows that in programming
Eq. (5) for a computer, the finite size of the arrays
55
VOLUME 62, NUMBER I JANUARY 1972
WILLIAM HADLEY RICHARDSON
150
140
130
12D)
110 [
100
90 i
80 I
70
60 I
50
40
30
20
10
0
0 1 2 3 4 5 6 7 8 9
Iterations
FIG. 1. Convergence of restoration toward
true values of undegraded image.
allows Eq. (5) to be rewritten
c Sk-i+i~k
Wi,,+x=WiV EVb
k-i b
E Sk-j+l'Wj. r
j-a
10
(6)
I I I I I I I I I
(2.2)
(2.1)
- a
(1.2)j
I I I I I I I I I
1.45
1.4
1.3
C
where a= (1, k-J+ 1).ax. b= (kjI)min, andc=i+J-1.
When Bayes's postulate is used for the initial estimate,
Eq. (6) becomes
c Sk-i+,Hk
wi,
1
= E
kiw b
Z Sk-j+1
J-a
(7)
for the first iteration. In Eq. (6), it appears that the
summation over k is a corrective factor on Wi,,. If
Wm.r were incorrect, the W.,, in the denominator of the
summation tends to correct the W-a,, in the first term of
the right side of the equation. For example, if in Eq. (6)
Wi,. were too large, the recurrences of Wn,, in the
denominator where a<m<b would tend to reduce the
value of Wm,r in calculating W.,,+,. This does not al-
ways occur, for the value of Wm.r may change in the
wrong sense initially to compensate for a relatively
large change in the value of another Wi,, in the
neighborhood.
1.2
1.1
1.0
0.9
0.8
0.7
0.6
0.55
Iterations
FIG. 2. Convergence of correction-factor
values (C) toward unity.
Equations (6) and (7) expanded to two dimensions
are
f f Hm nSm-i+I n-j+ (
m=,,, b d
W P WpqrSm-p+In-+
p-a q-d
and
e I Hm,nSm-i+I,n-j+1
Wi'j'I= E7 E -y (9
m-i n-j b d
at A' Sm-P+1,n-I +
p-a qsa
where
a= (1, m-K+1)max; b= (mI)min;
C = n12 Lt + 1)mx d1= (ntj)mjin;
e=i+K-1; f=j+L-1;
i= (1,l}; j=(1,J}.
K, L are the dimensions of Ski.
I, J are the dimensions of Wi j.
56
Vol. 62
ITERATIVE METHOD OF IMAGE RESTORATION
An example of the convergence of the process is
shown in a simple two-dimensional, noiseless system in
Fig. 1, where
2 10 12
H= 10 60 52 ; S=[2 4]
30 82 56
The undegraded image was
C 20 60]
100 140-
Figure 2 shows the trend toward unity of the correction
factors that gave the results for each iteration.
Figure 3 shows a restoration of two targets without
noise and Fig. 4 shows two targets with 0.1 random,
multiplicative noise added after degradation. The
degradation was by a uniform point spread function
five units square. In these figures, the pairs are in the
order (A) undegraded image, (B) degraded image,
(C) 10-iteration restoration, (D) 20 iterations, and
(E) 30 iterations. The specification of added noise is
defined by r in
H'= H(I+ rd),
A
B
C
D
E
(10)
FrI. 3. Restoration withi nonoise (A) Original image, (B) de-
graded image, (C) 10-iteration restoration, (D) 20 iterations, and
(E) 3C iterations.
110 A
B
C
E
Fic, 4. Restoration with 0.1 noise. (A) Original image, (B) de-
graded image, (C) 10-iteration restoration, (D) 20 iterations,
and (E) 30 iterations.
where ftT is the noisy value, H is the noise-iree value,
r is the fraction of added noise, and d is the random
deviate, normal (0,1). In this example, r is 0.1. The
two targets were designed with the idea of simplicity,
but including a certain amount of detail. They were
designed also to make it difficult to relate either de-
graded image with either undegraded image. The
approximating process appears successful because the
restorations are readily identifiable after only 10 itera-
ations. In the case of no noise, the procedure tends
toward a perfect restoration, whereas with noise the
restoration continues to improve.
Figure 5 shows restoration of the same noisy degraded
images by the Fourier-transform method using a least-
squares filter with estimated parameters to reduce the
effects of noise. The presence of background noise in
the restoration should be noted in contrast to the lack
of it in the iterative restorations. This could be a
detriment in more-complex images.
The effect of noise introduced after degradation was
investigated. One element of a noiseless degraded image
was doubled and the resulting image was restored. The
undegraded image wvas a square uniform field five points
on a side. The point spread function was a square
uniform field three points on a side, resulting in a
5 7
January 1972
WILLIAM HADLEY RICHARDSON
FIG. 5. Restoration with 0.1 noise by the Fourier-transform
method with estimated least-squares filter.
seven-by-seven degraded image. First, successive points
for noise addition were chosen on a diagonal of the
seven-by-seven degraded image, progressing point by
point from a corner to the center of the image. Next,
points were chosen on an axis, progressing from the
center of a side to the center of the image. Figures 6
and 7 show the results after 10 iterations of the restora-
tion process.
This process has been programmed and run exten-
sively on the CDC 3600 computer at the University of
California, San Diego, Computer Center. In pro-
gramming, the massive calculations indicated in Eq.
(8) were reduced by a very large factor by removing
redundant calculations.
The restorations shown in Figs. 3 and 4 averaged 7.4 s
per iteration.
CONCLUSIONS
Although no proof of convergence on a solution has
been devised, the process did converge in all cases for
which it was used. This seems reasonable because the
process is essentially self-correcting in the course of a
sequence of interdependent adjustments. In effect, the
.. '
I.'
.''.B
"I ','
FIG. 6. Restoration of originally uniform image (all values fo
Wij= 1.0). One value on diagonal of degraded image doubled.
(A) I1'(0,0)=2H(0,0), corner: maximum W=1.380, minimum
W=0.850; (B) H'(1,l)=2H((,1): maximum W=1.474, minimum
W=0.807; (C) 1I'(2,2)=211(2,2): maximum W= 1.494, minimum
W=0.819; (D) 11'(3,3)=211(3,3), center: maximum W=1.320,
minimum W=0.863.
.'. I
-. I
I' B
,' .
FIG. 7. Restoration of originally uniform image (all values of
Wi.i=1.0). One value on an axis of degraded image doubled.
(A) 1l'(3,0) = 211(3,0), middle of side: maximum W = 1.348, mini-
mum W=0.837; (B) If'(3,1)=2H1(3,1): maximum W=1.307,
minimum W=0.876; (C) 11'(3,2)=2H(3,2): maximum W=1.315,
minimum W=0.882; (D) 11'(3,3)=21I(3,3), center: maximum
W=1.320, minimum W=0.863.
process may be considered an iterative deconvolution
approximation.
In the examples of noise added at a single point in
the degraded image, a certain amount of disturbance
results in the restoration. However, this does not appear
to be out of control, as would often be the case with the
Fourier type of restoration. In the case of generally
distributed random noise, the disturbances tend to
compensate each other. On the other hand, the cases
in which noise was added at or near the edge of the
degraded image give warning of what may happen when
a part of a degraded image is processed. Here, the edges
of the degraded image may be contaminated by energy
coming from an area in the original image outside the
area of restoration. In practice, this results in a very
bright edge around the restoration, almost entirely
confined to the outer row or two. Effects of the ex-
traneous noise in the central part of the restoration are
very slight. This result can be countered by eliminating
the outer elements of the restored image after restora-
tion or by judicious tapering of values in the edges of
the degraded image.
The value of this process is that it can give intelligible
results in some cases where the Fourier process cannot,
although at higher cost than the Fourier process.
Equations (1)-(4), adapted with the notation of
Eqs. (6) and (7), indicate how this type of process may
be applied to some statistical uses of Bayes's theorem
when the a priori probabilities P(Wj) are not known.
ACKNOWLEDGMENTS
The author is grateful to J. L. Harris, Sr., for sug-
gesting this problem and affording the opportunity to
Vol. 62
ITERATIVE METHOD OF IMAGE RESTORATION
work on it. Dr. R. W. Preisendorfer has made many
valuable suggestions during the development of this
process. A. Shaules and D. P. MacAdam have acted as
helpful sounding boards during the investigation.
REFERENCES
* This research was supported by the Advanced Research
Projects Agency under Air Force Contract F08606-68-C-0017.
J. L. Harris, Sr., J. Opt. Soc. Am. 56, 569 (1966).
2 B. L. McGlamery, J. Opt. Soc. Am. 57, 293 (1967).
E. Parzen, Modern Probability Theory and Its Applications
(Wiley, New York, 1960), Ch. 3, Sec. 4, p. 119, Eq. 4.16.
4 The International Dictionary of Mathematics (Van Nostrand,
Princeton, N. J., 1960), p. 70, "Bayes's Theorem."
I R. E. Machol and P. Gray, Recent Developments in Information
and Decision Processes (Macmillan, New York, 1962), p. 175,
"Stable Estimation."
January 1972
59