Efficient curvature-based shape representation for similarity retrieval

The Curvature Scale Space (CSS) image is a multi-scale organisation of the inflection points of a closed planar curve as it is smoothed. It consists of several arch shape contours, each related to a concavity or a convexity of the curve. In our recent work, we have used the maxima of these contours to represent the boundary of objects in shape similarity retrieval. In our new approach, each segment of a shape is represented by the relevant maximum of the CSS image as well as the average curvature on the segment at certain level of scale. In this paper we explain how this new representation together with its matching algorithm improve the performance of our shape similarity retrieval system. To evaluate the proposed method, we created a small classified image database. We then measured the performance of the system on this database. The quantified results of this test provided supporting evidence for the performance superiority of the proposed method.