关键词: 序列保持, 首位区分, 哈希算法, 近邻检索

Abstract:

For traditional algorithms, the ambiguous ranking orders often cause an inferior ANN search performance. To solve this problem, the Ambiguous Ranking Perceptual Hashing（ARPH） is proposed. During the training process, the rule for distinguishing ambiguous ranking is defined in advance, and the hashing functions which satisfy the specify rule are learnt. As a result, the ranking orders of the data points which share the same Hamming distance to a query sample can be effectively distinguished without computing the bitwise weight and weighted Hamming distances. The objective function based on the ANN search performance evaluation metric of average precision is defined, and it belongs to the ordinal relation preserving restriction. Thus, the original Euclidean relative similarity relationship can be well preserved in the Hamming space. During the training process, the binary code, Hamming distances and judgment values are relaxed to continuous formulation. Finally, the gradient descent algorithm can be used to learn hashing functions. The ANN search comparative experiments on three image datasets show that the proposed ambiguous ranking perceptual hashing can achieve the best ANN search performance.

Key words: ordinal relation preserving, first position decision, hashing algorithm, approximate nearest neighbors search