Detecting objects in image collections using bipartite graph matching

Object-based image retrieval (OBIR) problem, in which the user is only interested in a fraction of the image, remains unsatisfactory, as it relies highly on accuracy. To address this problem, a novel method basing on bipartite graph matching is proposed in this paper. On the basis of the extraction of image features, we define a cost function according to the bipartite graph theory and minimize it by using the optimization technique to obtain an optimal map. Then, we calculate the mahalanobis distance to eliminate the mismatched points, since it takes into account the distribution of matched points. Finally, we apply the measure of coefficient of variation to evaluate the discrete degree and rerank the retrieved images. The experimental results on real video sequences and Caltech256 dataset demonstrate the effectiveness of our approach.