Hash Ranking With Weighted Asymmetric Distance for Image Search

Yuan Cao, Heng Qi, Jien Kato, Keqiu Li
2017 IEEE Transactions on Computational Imaging  
Image search can be viewed as a problem of large-scale Approximate Nearest Neighbor (ANN) search in image feature space. Hash ranking methods have been widely used for ANN search because of their two benefits: less memory usage and high search efficiency. Generally, the hash ranking methods face two problems: binary encoding and binary code ranking. This paper focuses on the latter. In existing work, the ranking of binary hash codes is usually implemented based on Hamming distance or asymmetric
more » ... distance. Hamming distance easily leads to confusing ranking when different candidate points share the same Hamming distance to the query point. Therefore, recent work prefers the asymmetric distance to Hamming distance. When computing asymmetric distance, it is necessary to give reasonable query-independent values. These values are usually approximated by average values of sample candidate points in existing methods. However, when the distribution of candidate points is not uniform, average values are meaningless, leading to wrong ranking results. To address this problem, we propose two kinds of weighted asymmetric distance algorithms, namely, the otsu threshold based algorithm (WoRank) and the score calculation based algorithm (WsRank) in this paper. The processes of these two proposed algorithms are similar, consisting of two steps. In the first step, we compute the query-independent values on each bit in accordance with corresponding distribution of candidate points to reduce the approximation error. In the second step, we compute bitwise weights in consideration of each bit's discriminative power to further improve the retrieval accuracy. The differences between WoRank and WsRank are the computation methods of query-independent values and bitwise weights. To evaluate the proposed algorithms, we conduct a large number of experiments on four well-known datasets, namely, SIFT, CIFAR-10, MNIST and NUS-WIDE. The results show that the proposed algorithms can achieve up to 22% performance gains over Hamming distance based ranking and 13% over the existing asymmetric distance based ranking. We also find WoRank is suitable for feature dataset (SIFT), while WsRank is suitable for image datasets.
doi:10.1109/tci.2017.2736980 fatcat:rnqxsrlqu5c6xnlrgcijmj5d3e