ANALYSIS OF RELIABILITY AND IMPACT FACTORS OF MUTUAL INFORMATION
SIMILARITY CRITERION FOR REMOTE SENSING IMAGERY TEMPLATE
MATCHING
H.L. Wang 9 ’, R. An b , * Q. Zhang b , C.Y. Chen b
a College of Geography and Sea Sciences, Nanjing University, Nanjing 210093;
b Geo-informatics Department of College of Hydrology and Water Resources, Hohai University, Nanjing 210098.
KEY WORDS: Matching, Registration, Reliability, Performance, Impact Analysis, Multi-sensor
ABSTRACT:
Reliability of similarity measure is a key factor for successful image matching. In this paper, mutual information is introduced to
template image matching as a similarity measure. Its ability of adapting to different scene, overcoming the grey reversal, and factors
which influence on the successful matching of mutual information method are discussed. These factors include signal-noise ratio,
information content and self-similarity. The test data are remote sensing images captured by different sensors, with different spatial
resolution and at different seasons. Some typical scenes such as city, town, village, river, stream, and field land are used to test its
ability of adapting to different scenes and the ability to overcome the grey reversal. The experiments manifest that there is no strong
relationship between success rate and the amount of signal-noise ratio or information content for the mutual information matching
method. The successful matching based on mutual information is mainly affected by whether there are similar pattern areas in the
reference image or not. The experiments also indicate that when the scene has large non-linear change of grey, the success rate of
mutual information matching method is much greater than that of cross-correlation method.
1. INTRODUCTION
Template matching is essential in many image analysis and
computer vision tasks. Reliability of similarity measure is a key
factor for successful template matching. In conventional template
matching methods, cross-correlation methods are often used
because of their strong suitability to different scene, simple
arithmetic and easily parallel calculation. However, when a big
grey aberrance (such as grey reversal), geometrical deformation
(such as rotation) and random noise exist in images to be matched,
the success rate of matching is low and matching usually fails.
Therefore, it is necessary to study more reliable similarity measure
to boost up the ability to overcome intensity aberration and random
noise in order to improve the success rate of matching (Brown,
1992; Zitova et al., 2003;Su et al., 2000).
In this paper, mutual information is introduced to template image
matching as a similarity measure. Its ability of adapting to different
scene, overcoming the grey reversal, and factors which influence
on the successful matching of mutual information method are
discussed. Mutual information is a fundamental concept in the
information theory, and it is the measure of the statistic relativity
for two stochastic variables. When images with the same structure
matched best, the mutual information of corresponding pixels is the
biggest in principle (Maes, et al., 1997; Studholme, et al., 1999;
Arlene, et al., 2003). Because mutual information similarity
criterion doesn’t need any assumption of pixel value relationship
between the images to be matched, and any segmentation and pre
processing are not required before matching too, it has been widely
applied in the matching of various kinds of images such as
medicine images (Josien, et al., 2003). Whereas, in template
matching, there are few articles discussed about it so far as I know,
so it is a valuable topic to be discussed farther.
The factors, which affect the successful matching of mutual
information method, are also analysed. These factors include
signal-noise ratio, information content and self-similarity. The
relationship between these factors and success rate of matching are
presented. The success rate of mutual information and cross
correlation based methods are compared. Some beneficial
conclusions are drawn.
The rest of the paper is organized as follows. The basic principle of
mutual information is described in Section II. The test data and
matching method are described in Section III. Reliability and
impact factors of mutual information similarity criterion are
discussed in detail in section IV. Finally, some conclusions are
drawn in Section V.
2. MUTUAL INFORMATION SIMILARITY METRIC
Mutual information (MI) is a concept developed from information
theory. It indicates how much information one random variable
tells about another. The MI registration criterion can be thought of
as a measure of how well one image explains the other. It is applied
to measure the statistical dependence between image intensities of
corresponding pixels in both images, which is assumed to be
maximal if the images are geometrically aligned; therefore, it can
be regarded as the similarity measure in image matching (Maes, et
al. 1997; Viola, et al.1997; Arlene, et al. 2003; Josien, et al. 2003).
It has been widely applied in the matching of various kinds of
images such as medicine images that obtained with different mode,
and also has been used in remote sensing imagery registration
recently (Chen,et al. 2003).
The mutual information is denoted by information entropy as
follows:
‘Corresponding author. Tel.: +86 025 83787578; E-mail addresses: anrunj@yahoo.com.cn; anrunj@163.com.
1129