Learning the shape manifold to improve object recognition

Abstract

This paper presents an approach for object recognition and shape retrieval in binary images. Object description is accomplished by contour and region-based solutions. Fusion of two categories of shape descriptors causes a considerable performance improvement in retrieval. In order to improve the recognition rate, it is important to extract the intrinsic object feature vectors in such a way that the geometrical distance between two such vectors matches the semantic relation between the two objects from which they were extracted. The main contribution in this paper is learning how to map the samples in high-dimensional observation space into the new manifold space so that the geometrically closer vectors belong to near semantics. Feature extraction is done by Isomap that belongs to a non-linear feature extraction algorithm. In the proposed method, the shortest path in the dissimilarity graph leads to less value between the samples in the same category by denoising the edge values. Experiments show that the geometrical distance between the samples on the manifold space are more compatible to the semantic distance of them. The proposed method has been compared to some well-known approaches by a variety of shape databases which includes MPEG-7 Part B, Kimia silhouettes, and Fish. The results show the effectiveness and validity of the method.