Performance analysis of various local and global shape descriptors for image retrieval

In this paper, various prominent local and global descriptors are evaluated against each other for analyzing their performance on shape-based image retrieval. Local descriptors include Fourier descriptors, Weber's local descriptor, local binary patterns, and local ternary patterns. The prominent global descriptors include moment invariants, generic Fourier descriptor (GFD), angular radial transform (ART), wavelet moments (WM), and Zernike moment descriptor (ZMD). In addition, a novel local descriptor is proposed based on the histograms of circular arcs and linear edges, which are detected by means of Hough transform. The proposed local descriptor provides features, which are invariant to geometric transformations and are robust to noise as compared to some existing prominent local descriptors. We also propose an improvement in the performance of global descriptors GFD, ART, WM, and ZMD by taking advantage of the phase information in the comparison process along with their magnitude. Subsequently, the local and global descriptors with the best image-retrieval performances are combined to design an effective retrieval system, which further enhances the retrieval performance substantially. All descriptors are analyzed in terms of six principles set by MPEG-7. Detailed experiments are performed on standard benchmark image databases along with their rotation-invariance and noise test. The results of experiments reveal that the proposed fusion of local and global descriptors outperforms other major descriptors.