Statistical similarity measures in image retrieval systems with categorization & block based partition

This paper presents a novel approach of similarity matching in image retrieval based on the distribution of joint feature vectors of color and texture features. Mean vectors and covariance matrices are computed from feature distributions of training samples with known categories and from individual images with varying partitions on the assumption that, distributions are multivariate Gaussian. Statistical distance measures utilize these parameters in similarity matching functions to minimize the probability of retrieval error. For category specific retrieval, a multi-class support vector machine (SVM) is trained on the samples to predict the categories of query and database images. Based on the online prediction, precompiled category specific statistical parameters are utilized in similarity measure functions. For partition specific retrieval, individual images are partitioned into non-overlapping blocks of different sizes and a joint feature vector of color and texture features are extracted from each block to generate the distribution and estimate the parameters. Experimental results on a generic image database with ground truth are reported. Performances of two statistical distance measures, namely Bhattacharyya and Mahalanobis are evaluated and compared with Euclidian distance measure, which show the effectiveness of the proposed technique.