Adaptive non-cartesian networks for vision

Abstract

We address the problem of locating and extracting frame curves on interesting image areas. Reference frames, focus of attention, bounding contours of shapes, axis of inertia, centers of masses and other mid-level visual structures, can be used to guide other mid-level visual tasks or to lead subsequent high level processing like recognition, indexing or image retrieval. Frame curves, are useful to tackle non-rigid object recognition problems because these have fuzzy boundaries. Where is the boundary of a cloud, oak leave or a leopard? we present a perceptual organization approach based on dynamic programming and adaptive non-cartesian networks, a new kind of networks which are based on placing processor lines using a distribution function adapted to the image array. We present a novel computational framework to extract frame curves directly on the image and several experiments on real images.