Create efficient visual codebook based on weighted mRMR for object categorization

The bag-of-words approach is gained much research in object categorization. Creating visual codebook is an important problem in object categorization. The non-informative codeword will increase the vocabulary size which brings more computation cost, and cannot improve the classification performance. We first define a weighted minimal-redundancy-maximal-relevance criterion (mRMR) which is an extension of basic mRMR. And we propose an iterative method to select efficient visual words based on weighted mRMR in backward way. We first get the initial set of codewords through k-means cluster, then we use the proposed method to select the most discriminative subset of codewords which are used to compute histograms. We perform experimental comparison of our algorithm and basic BOV on Caltech database. The experimental results proved that the proposed algorithm can achieve good performance and lower computation cost with smaller size of vocabulary.