mean and variance transformation to each block im- 
; age. The new mean and variance of each block im- 
age are determined by a rule base. (3) Recombine 
s the enhanced block images together into a high qual- 
i ity output image free of border effect, based on a 
Gaussian image overlay scheme. 
The purpose of localized enhancement is to im- 
prove the local brightness of an image. The differ- 
; ence between local and entire image enhancement is 
; that when local image enhancement is applied, we 
; must take into consideration some a priori know- 
ledge,e.g. if the mean and variance of a block image 
1 are small, the block image must be very dark re- 
gion of input image, and probably is the back- 
f ground of a scene, then it doesn' t have to be en- 
hanced. In order that both local and entire image 
  
  
  
  
  
  
  
  
  
  
Fig.3 
(b) 
enhancement use a rule base for new/old mean and 
1 variance transformation, a judgement rule base is 
(c) 
necessary, which is only used for local image 
enhancement to judge whether the mean and vari- 
. ance transformation are applied to a block image or 
; not. If the mean and varicance range are divided in- 
to 5 regions respectively as shown in Fig. 4, then 
we can acquire 14 rules. For example, 
judgement rule 1: 
IF: (1)the mean is vs, (2) the variance is s 
THEH: keep the block image unchanged 
In order to eliminate the border effect resulted 
from localized enhancement between block images, a 
Gaussian image overlay technique is proposed. 
  
  
  
  
i 4 i + i d 
2 0 10 80 180 240 255 
ph — —M — he —| 
| ENS 8 |. m vw 
(a) Semantics of Mean 
; 
à: = + + b + = 
; 0 6 20 4000 6000 10000 
Mp ii P >i Ri — 
P NS ^! S m 1 vl : 
; 
Fig. 4 (b) Semantics of Variance 
The so-called Gaussian overlay image technique is 
to acquire overlay block subimages based on the 
667 
overlay block structure as shown in Fig. S. If an 
image is divided into 25 subimages, 25 overlay-block 
images can be acquired. Each overlay-block image is 
the union of corresponding block subimage with its 
eight-neighboring block subimages in Fig. 3(c)If the 
subimage is situated at the corner or on the border 
of an image, the overlay-block image is the union of 
corresponding block subimage with its three or five 
neighboring block subimages. Then, 2-dimensional 
Gaussian spatial weighting function is applied to 
compute again the new/old mean and variance for 
each pixel in the mean and variance transformation. 
The Gaussian spatial weigting funtion is 
Px(ij)- Exp(-[G— ig)? * (j5.)1]/2o2) (10) 
where(i,j) is the coodinate of pixel, (ig.jk) is the 
center point coodinate of block image k in over- 
lay-block image, s? is variance. The center point 
and variance are illustrated in Fig. 3(a) and Fig.3(b). 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Fig. 5 i i 
EXPERIMENTAL RESULTS AND 
CONCLUSIONS 
Fig. 6(a) is a 480-by-512 pixel image with a grey 
level range[0,255], of which the grey levels so heavi- 
ly on low and high levels that many mini scenes 
cannot be distinguished. We have enhanced the origi- 
nal image applying tranditional image enhancement 
algorithms, e.g., histogram equalization, logarithm 
stretch, and the 
rule-based localized image enhancement technique 
presented in this paper, respectively. For the limita- 
tion of this paper here is only given the experi- 
mental result of histogram equalization. The results 
demonstrate that the best effect is acquired by ap- 
plying the rule-based localized image enhancement 
technique, and the border effect is eliminated com- 
stretch and  piecewise linear 
pletely. 
The technique presented in this paper carries out 
the task of automated selecting parameter quite well 
in image contrast enhancement. Because the human 
being knowledge about image enhancement technique 
and the expert evalution of enhanced image have 
been summarized to rules, which can be trained and 
modified. The benefits of the rule-based localized im- 
age enhancement techique are of high efficiency, sta- 
ble performance, strong adaptability and best effect.