On a relaxation-labeling algorithm for real-time contour-based image similarity retrieval

摘要

In this paper, we propose a relaxation-labeling algorithm for real-time contour-based image similarity retrieval that treats the matching between two images as a consistent labeling problem. To satisfy real-time response, our algorithm works by reducing the size of the labeling problem, thus decreasing the processing required. This is accomplished by adding compatibility constraints on contour segments between the images to reduce the size of the relational network and the order of the compatibility coefficient matrix. Particularly, a relatively strong type constraint based on approximating contour segments by straight line, arc, and smooth curve is introduced. A distance metric, defined using the negative of an objective function maximized by the relaxation labeling processes, is used in computing the similarity ranking. Experiments are conducted on 700 trademark images from the Japan Patent Office for evaluation.