Evaluation of Shape Descriptors for Shape-Based Image Retrieval

This paper presents a comparative study between scale, rotation and translation invariant descriptors for shape representation and retrieval. Since shape is one of the most widely used image feature exploited in content-based image retrieval systems, we studied for each descriptor, the number of coefficients needed for indexing and their retrieval performance. Specifically, we studied Fourier, curvature scale space, angular radial transform, and image moment descriptors for shape representation. The four shape descriptors are evaluated against each other using the standard methodology and the two most appropriate and available databases. Results showed that moment descriptors present the best performance in terms of shape representation quality while angular radial transform presents the lowest descriptor size.