Fast near duplicate detection for personal image collections

Due to the rapid growth in personal image collections, there is increasing interest on automatic detection of near duplicates. In this paper, we propose a novel fast near duplicate detection framework that takes advantages of heterogeneous features like EXIF data, global image histogram and local features. To improve the accuracy of local feature matching, we have developed a structure matching algorithm that takes into account of a local feature's neighborhood which can effectively reject mismatches. In addition, we developed a computation-sensitive cascade framework to combine stage classifiers trained on different feature spaces with different computational cost. This method can quickly accept easily identified duplicates using only cheap features without the need to extract more sophisticate but expensive ones. Compared with existing approaches, our experiments show very promising results using our new approach in terms of both efficiency and effectiveness.