Probabilistic Similarity Measures in Image Databases with SVM Based Categorization and Relevance Feedback

Abstract

This paper demonstrates an approach to image retrieval by classifying images into different semantic categories and using probabilistic similarity measures. To reduce the semantic-gap based on low-level features, a relevance feedback mechanism is also added, which refines the query parameters to adjust the matching functions. First and second order statistical parameters (mean and covariance matrix) are pre-computed from the feature distributions of predefined categories on multivariate Gaussian assumption. Statistical similarity measure functions utilize these category specific parameters based on the online prediction of a multi-class support vector machine classifier. In relevance feedback, user selected positive or relevant images are used for calculating new query point and updating statistical parameters in each iteration. Whereas, most prominent relevant and non-relevant category specific information are utilized to modify the ranking of the final retrieved images. Experimental results on a generic image database with ground-truth or known categories are reported. Performances of several probabilistic distance measures are evaluated, which show the effectiveness of the proposed technique.

This work was partially supported by grants from NSERC and ENCS Research Support.