A new training-based approach for robust thresholding

This paper presents a training-based approach for threshold selection in digitized images. The work is motivated by the difficulties of adapting existing algorithm to particular computer vision applications. The algorithm proposed is based on a learning process that extracts heuristic knowledge from training images and incorporates it in in fuzzy systems to be applied in the supervision of a fuzzy-multiresolution threshold selection method. This threshold method selects the set of intensity values of the object pixels by analyzing the multi-scale decompositions of the image histogram under the supervision of fuzzy systems that contain the knowledge incorporated during the learning process. The methodology allows easy adaptation to specific computer-vision applications. The particularization to one application is presented to show its performance.