A compact feature representation and feature re-weighting in content-based image retrieval

Proper selection of features and their representation are essential for a good Content-Based Image Retrieval (CBIR) system. In this paper, we propose a compact feature representation based on the elements of Colour Co-occurrence Matrices (CCM) in Hue, Saturation, Value (HSV) colour space. To represent each image in the database, we constructed a feature vector considering all diagonal elements and Sum-Average [1] of all non-diagonal elements of the CCM. We demonstrate that our feature representation is superior in terms of system accuracy and online computation time which are attributed to dimension reduction. We present experimental results with a database of 2000 images from 10 categories. With our method, we achieved up to 3 improvement in precision. With relevance feedback, we obtained a further 12 improvement as opposed to 9 with original higher dimension.