Optimum Template Selection for Image Registration Using ICMM Adnan A. Y. Mustafa Department of Mechanical and Industrial Engineering Kuwait University P. O. Box 5969 - Safat Kuwait 13060 symymus@kuc01.kuniv.edu.kw Abstract In our previous work on image registration we developed a novel image registration method, called the Intensity Combinatorial Minimization Method (ICMM), that has many appealing features. Two important features of this method is that ICMM is computationally efficient and has the unique feature of being invariant to the image processed by an injective function. In this paper we extend the use of ICMM for template matching. The extraction of an optimum template is investigated. Optimization of both template location and template size are addressed. We introduce the Intensity Variation Number, which is an image information measure that is strongly related to entropy. We show that optimization is a function of the Intensity Variation Number. Results of tests conducted on real images with noise are presented that support our theories. 1 Introduction Image registration and template matching is an area of active research in computer and machine vision. The importance of registration and template matching are not limited to this field but are also important in many other fields because of their frequent occurrences. They arise in the domain of object recognition [1], multi-spectral image analysis [2], aerial image analysis [3], meteorology [4], medical imagery [5], and many more [6]. The main goal of the template matching process (which is at the core of image registration) is to find the translational, rotational and scaling offsets between the template and the image. In our previous work on image registration [7] [8], we had developed an efficient area-based image registration method, called the Intensity Combinatorial Minimization Method (ICMM), that has many appealing features. In addition to being computationally efficient, where only simple calculations are required, the method has the unique feature of being invariant to the image (or the template) processed by an injective (one-to-one mapping) function. Tests conducted with this method produced good overall registration results indicating that this method is more efficient and robust than other traditional registration methods. In this paper we investigate the extraction of an optimum template. Optimization of both template location and template size are addressed. We will show BMVC 1998 doi:10.5244/C.12.81