Segmentation approach by learning: different image applications

Abstract

We present a new segmentation approach by thresholding based on learning strategy. This strategy is based only on one image and its ideal thresholded version. A decision matrix is generated from each pixel and each gray level. At the moment of new image segmentation, the best solution for each pixel is evaluated by means of K nearest neighbors in the decision matrix. Comparative tests were performed on signature, fingerprint and magnetic resonance images.