Two-dimensional object recognition using simulated annealing

Abstract

The following problem, which is intrinsic to computer vision, is considered in the paper How to recognize, in a given scene (available in the form of a digital image), an object that is one among many possible items found in a certain knowledge (or model) base? The object in the scene could have been subjected to the following 'distortions': scale and orientation changes, and a shifting of position (with respect to the corresponding model in the knowledgebase). It is assamed that the given image is segmented and the objects are not subject to occlusion, I.e., each object in the scene is isolated from the others, and can be located independent of the rest by examining the appropriate regions of the Image. We present a new method for solving this problem. In contrast with the results of the literature, the objects' line segments, arcs and the like, are not stored in the knowledge-base. The price paid for this is- that the matching strategy becomes more complicated. The novelty in the proposed method, however, is believed to be in the reformulation of the recognition problem as one of minimizing an error function, which, being non-convex, cannot be handled by standard optimization techniques. Therefore, simulated annealing is employed to find its global minimum. Nevertheless, this method is not recommended for real-time mantpulation of robotic arms controlled by a vision system.