ïng 2008 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1131 
>0 (3) 
V) (4) 
distribution of 
ility. For grey 
îy histograms, 
)int histogram. 
ne et al (1999): 
(5) 
nsitivity of MI 
ir experiments 
significantly 
f view. 
METHOD 
5 captured by 
lifferent spatial 
n) acquired on 
n are regarded 
Corresponding 
1 as the source 
date matching, 
meter. Several 
e obtained by 
cropping the source IRS-C image with an interval of 20 pixels 
apart. Images to be matched cover different landscape, such as city, 
village, river, farmland, etc. and they have different characters of 
signal -noise ratio, information content, and self-similarity pattern. 
MI is estimated by grey histogram, and NMI is used to eliminate 
the influence of the image overlap. Since the size of reference and 
input images are small, when the number of grey levels in each im 
age is high (e.g., 256 grey levels), the statistical power of the prob 
ability distribution estimation by mutual information will be re 
duced (Knops et al. 2006). Therefore, image grey-level reduction is 
needed. In the paper’s test, image grey-level is set as 16. The im 
plementing process of MI matching is as fellows: 
(1) Searching input image pixel by pixel and probability of 
every grey-level is estimated. 
(2) Moving the input image on the reference image pixel by 
pixel, and corresponding sub-image with the same size of the input 
image is cropped from the reference image. The probability of 
every grey-level is estimated for every sub-image. 
(3) Calculating the probability of the grey pairs presented in 
corresponding position of the images to be matched. The value of 
normalized mutual information is obtained in terms of information 
entropy formula (5). 
The position, on which the NMI is the greatest, is regarded as the 
correct matching position for the input image in the reference 
image. 
4. IMPACT FACTORS OF MUTUAL INFORMATION 
SIMILARITY CRITERION AND ITS VALIDITY 
Whether image matching is successful or not depends to some 
degree on the characteristic of images except for the matching 
method. Factors of image quality, which influence on the 
performance of the area-based matching, mainly include signal- 
noise ratio, self-similar pattern, the number of independent pixels, 
square difference. The signal-noise ratio often used to forecast 
whether the matching is successful or not by area-based matching 
methods (Su, et al., 2000). The information content is the main 
factor, which impacts the success of matching by feature - based 
methods (An, et al., 2005). All methods may be failure if some 
self-similar pattern areas present in the reference images. This 
paper pays mainly attention to how the signal-noise ratio, 
information content and self-similar pattern influence on the 
matching validity based on mutual information method. 
4.1 The Signal-Noise Ratio 
The signal-noise ratio is an important parameter for matching 
performance, which is mainly determined by weather status when 
the image is captured, the noise degree of sensor and the equipment 
used for digitising image. The greater the signal-noise ratio is, the 
more favorable it is for matching. Thus, the signal-noise ratio is an 
important factor to estimate image quality, especially in real time 
image matching for navigation and position estimation (Ma, et al., 
2001). It is applied to describe various errors of input image and 
reference image, and it is a more important rule for forecasting 
performance of matching (Du, et al., 2003). In scene matching, it 
implies the grey difference between corresponding pixel for input 
image and reference image. The definition of signal-noise ratio is 
as follows: 
SNR = - = - 
N 
VAR(ref) 
VAR 
ref - rel 
~ * VAR(ref) 
(6) 
VAR(rel) 
Where VAR(*) denotes the calculation of variance of image (see 
equation (7)); ref is the reference image processed, and its each 
pixel grey value is obtained by subtracting average grey value of 
ref image from corresponding pixel grey value in ref image; rel 
is the corresponding input image, which is also processed by the 
same method as ref image. Suppose the size of image / is RxC , 
/' 
is the average grey value image of / , then, 
VAR(I) = 
1 
RxC 
R-1 C-l 
/=0 y=0 
(7) 
Rate of successful matching 
ling 
rom 
0.32 0.34 0.36 0.38 0.40 0.42 0.44 0.46 0.48 0.50 
Signal-noise ratio 
Figure 3. Successful matching rate of images whose signal-noise ratio is between 0.3-0.5