In this paper, a novel feature descriptor named ''correlated microstructure descriptor (CMSD)'' is proposed for image retrieval. CMSD represents high level semantics by identifying microstructures via establishing correlations between texture orientation, color, and intensity features. The proposed CMSD follows a two-staged approach for feature extraction and their integration. Color information is extracted by exclusively quantizing each component of HSV color space. Richer edge orientation information is extracted by using the multi-directional Sobel operator. Local contrast information is obtained by quantizing V component of HSV. Correlated microstructures are then identified by correlations established on the basis of proximity and continuation relations. The identified microstructures are then mapped to color, texture orientation, and intensity features exclusively to obtain micro-color, micro-orientation, and micro-intensity maps, respectively. Proposed three concatenated micro-maps are represented through a 2-D histogram, incorporate information regarding semantic and spatial contents of an image. The experiments are performed on standard datasets, i.e., Corel 1k, Corel 5k, and Corel 10k. The results evaluation depicts the outstanding performance of proposed CMSD as compared to state-of-the-art methods including TCM, GLCM, MTH, MSD, MTSD, and SED.
INDEX TERMSContent based image retrieval, correlated low-level visual features, correlated primary visual features, micro-maps, microstructures.