Kernel fitting for image segmentation

Previously, a classifier called Kernel-based Nonlinear Representor (KNR) was proposed for pattern classification. In this paper KNR is changed to curve fitting for image segmentation applications. For each gray level, a curve is estimated by KNR and separated from that of a higher gray level by a threshold obtained from Newman-Pearson criterion. The thresholds are then merged into a few representative ones, with an ideal high-pass filtering approach, for image segmentation. Feasibility of the presented method in image segmentation is illustrated by some experimental results.