A Novel Method to Construct Visual Vocabulary Based on Maximization of Mutual Information

Bag of visual words model has shown interesting capabilities of object category recognition. Considering that there is different discriminative ability of each visual word for object categorization, we propose a novel method based on maximization of mutual information to measure the discriminative ability of words and then give a visual vocabulary optimization algorithm based on more discriminative words selection. Experiments showed that the method presented in the paper not only preserves more discriminative information and gives a further 7.9% classification performance increase over standard vocabulary, but also result in dimension reduction in large scale and classification efficiency increase. The experiments were performed on a 7-class object database containing 1776 images.