Rolling penetrate descriptor for shape-based image retrieval and object recognition

This paper proposes a new approach called rolling penetrate descriptor for shape description. It combines the advantage of the contour-based and the region-based methods, and provides an unified scheme to handle various shapes. The main process of the proposed method is to use a set of scanning lines that rotate around the shape centroid to collect information. During the rotating process, three feature functions are computed to reveal the inner structures of the candidate shape. The proposed method is very flexible and can be adapted for certain applications, while the scanning process serves as a framework. The rolling penetrate descriptor method is tested on several data sets with variations including common geometrical transform, noise, distortion and occlusion. Experimental results demonstrate that the proposed approach has strong capability in handling a variety of shapes.