BVLCs and BDIPs in Covariance Descriptor with a Statistical Measure of Divergence for Image Retrieval

Abstract

This paper proposes BVLC (block variation off local correlation coefficients) and BDIP (block difference of inverse probabilities) into covariance descriptors for image recognition and retrieval. The images in RGB is partitioned to R,G and B then BVLCs and BDIPs are computed for each channel then BVLCs and BDIPs in covariance among R,G and B is computed. The proposed BVLCs and BDIPs in covariance descriptor is extremely precious to reflect the degree of linear association among identified textures, edges and valleys. Degree of similarity among query and target images is analyzed using statistical measures of divergence namely Chernoff and Bhattacharya and results shown that Chernoff is noticeably outperform Bhattacharya measure. Comprehensive experiments on Corel- 1k, Corel- 5k and Corel-10k databases illustrates that proposed BVLCs and BDIPs in covariance descriptor is satisfactory and can attains significantly enhanced retrieval accuracies than traditional techniques.