Definition of a robust measurement of similarity for the localization of small shapes in scenes

Abstract

After analyzing the problems related to finding a small shape in a scene, we introduce a nonparametric similarity measure based on the Kolmogorov-Smirnov statistic, which proves to be robust for template-matching problems where a target of binary characteristics is to be located inside a gray-scale image. We show that the Kolmogorov-Smirnov statistic gives the optimum thresholding level for the image and may be computed without actual thresholding of the image. Some interesting properties of the proposed similarity measure are exposed and compared to the corresponding properties of normalized correlation.