Implicit model-oriented optimal thresholding using the Komolgorov-Smirnov similarity measure

We analyze the problem of local thresholding of a scene under the constraint of the geometric model of the target to be located, in the scope of the locating search process of an essentially-binary target in a gray-level scene. An optimal threshold is obtained which maximizes the fitting of the thresholded image to the target model template in the binary domain. As a result of this maximizing process the Kolmogorov-Smirnov similarity measure is obtained, which allows target spatial location in the scene with no need of explicit thresholding of the image, and avoids the high computing cost associated to gray-level domain similarity measures, such as normalized correlation. When used as a similarity measure, the nonparametric characteristic of the Kolmogorov-Smirnov statistic yields invariance and normalizing properties, which are shown to improve the properties of commonly-used gray-level domain similarity measures