Texture frame curves and regions of attention using adaptive non-Cartesian networks

摘要

This paper addresses the problem of locating and extracting frame curves in images with complex backgrounds. Frame curves are virtual image curves which lie along `the center' of texture boundaries. They can be used to carry out other visual tasks or to lead to subsequent processing such as recognition, indexing or image retrieval. We present an approach based on dynamic programming and adaptive non-cartesian networks (ANCN). ANCN are based on line placement using a distribution function adapted to the image texture. We present a computational scheme to extract frame curves directly from the image and we provide the results of several experiments on real images.