The contrast enhancement with mean and variance 
thansformation has following advantages: 1. It is a 
linear transformation. 2. The restraint role of a few 
limit grey levels can be eliminated so that the con- 
trast enhance effect can be improved as a whole. 
3. The transformation effect depends on the selec- 
tion of new mean and variance mainly, and the rela- 
tion between the new/old mean and variance is 
invariant basically, therefore the kind of images can 
be processed effectively if only the regularity men- 
tioned above has been found. 
ESTABLISHMENT AND MODIFICATION 
OF A RULE BASE 
The function of rules is to acquire the expected 
new mean and variance on the basis of the image 
fact bases. The rules are invariant with input 
image, i. e., independent of input image. The 
establishment and modification of rules are based on 
man' s knowledge and the analysis of processing re- 
sults for various images. 
As is well known, if the mean of a image with 
grey range[0,255] is nearly equal to the center value 
128 and the variance is large, the image must be 
very clear. Therefore, it is unnecessary for the image 
to be enhanced, i.e,the mean and variance of the 
image are kept unvaried. Otherwise, the mean of a 
source image should be transformed nearly to 128, 
and the variance be enlarged. Based on the above 
idea, a rule base can be established. 
The mean and variance range is divided into 7 re- 
gions respectively, which are expressed with 
semantics as shown in Fig.2. Then, 49 rules in all 
are generated based on the different combination of 
the means and variances. The format of a rule is 
^IF-" THEN -- " 
enhancement rule 10: 
IF: (1) Ms is vs, (2) Vs is s 
THEN: (1) M,=M;:1.5+30, (2) V,= V, 3 + 1500 
The semantics of vs and s are shown in Fig. 2. 
Now, the frame of a rule base has been formed. 
For example, 
195 225 
  
0 30 60 90 165 255 
! es el | 
E t 
S — 
m 
(a) Semantic Definitions of Mean 
L 1 1 i 1 4 i 4 
0 100 1000 2000 3000 6000 8000 10000 
— pw 
+=" br 
m E 
Fig .2 (b) Semantic Definition of Variance 
In practice, the best enhancement effect may not 
be sometimes achieved by the mean and variance 
666 
transformation based on the rules above. Therefore, 
it is necessary to modify the rules according to the 
analysis of image processing results. 
In the rule base, the transformation relation be- 
tween the mean and variance of output image and 
those of source image is 
Mo = Ms.a + C1 | (7) 
Vo = Vs-b+Ca | 
where C,, C, are constants, a, b are modifiable 
parameters. The procedure of solving the modifiable 
parameters according to the best result acquired by 
interactive method is given as follows: 
(1) Apply the mean and variance transformation 
to an image based on the initial 
(2) Display the enhanced image and the curve of 
Fig.l (including the values of Isi and Is2) 
(3) Answer the question: "Are you satisfied with 
the enhanced image? " If yes, goto(8) Otherwise, 
(4) Indication: "If to increase (or decrease) the satu- 
ration of low grey levels, increase (or decrease) Is: . 
If to increase(or decrease) the saturation of high 
grey levels, decrease(or increase) Is2. Input Is: and 
rule base. 
  
  
Is2 .” 
(S) Solve Mo and Vo with Isi and Is2 input: 
) : 
Mo7(123—) Ms-Isi) | 
Isz—Isi | (8) 
er 255-2 ( 
Net ioi vs | 
(6) Apply the mean and variance transformation 
based on the modified Mo and Vo. 
(7) Goto (2) 
(8) Answer the question: " Modify the rule base or 
not?" If no, goto (9). Otherwise, modify the 
parameters a and b of corresponding rule in rule 
base as follows: 
  
aimi z25- Mrs e 3 
p 5 2 © [ 
Is2— Isi Vs | 
(9) End. 
RULE-BASED LOCALIZED IMAGE 
ENHANCEMENT TECHNIQUE 
When an input image has strong spatially depen- 
dent variation in scene illumination, the mean and 
variance transformation illustrated above based on 
the global information over an entire image can not 
obtain high-quality output images. For this reason, 
a powerful and elegant localized image enhancement 
technique is presented in this paper. 
The scheme of the localized image enhancement 
technique is illustrated as follows: (1) Divide an 
input image into several block subimages, e. g., 25 
block subimages as shown in Fig.3(c) (2)Apply the