Fast Keypoint Reduction for Image Retrieval by Accelerated Diverse Density Computation

Through many researches on CBIR (Content Based Image Retrieval), local image features defined on keypoints are proved to be effective for occlusion-robust image retrieval. However, not all features in an image contribute correct image retrieval, i.e. some are unstable and some commonly appear in other images. For representing the importance of the image features, Diverse Density (DD) based index-feature reduction has been proposed[13], which reduces the elapsed time, as well as the number of erroneous retrievals. The only drawback of this method is low computational efficiency: it requires distance computations between all possible combinations of local features for computing RBF kernel. Because of this problem, the DD based keypoint reduction cannot be applied to a large scale image database. This paper presents an acceleration method of DD computation by applying Range Nearest Neighbor (RNN) search for gathering the proximal features to a given point in the feature space. This is because the distant points from the point in the feature space are believed to have negligible influence to the DD value at that point. We proposed two types of acceleration methods, one has a single index structure for the radius search and the other has multiple indices defined on single images. We applied our method to Nister's data set and achieved 520-times speedup over the original method. Also, we found that this approximation also improves the image-retrieval accuracy.