Exponential Hashing with Different Penalty for Hamming Space Retrieval

Abstract

Hamming space retrieval enables efficient constant-time search through hash table lookups constructed by hash codes, where in response to each query, all data points within a small given Hamming radius are returned as relevant data. However, in Hamming space retrieval, the search performance of the existed hashing schemes based on linear scan dropped when the length of the hash codes increases. The reason is that the Hamming space becomes very sparse and it is difficult to pull the similar data into the Hamming ball and to push the dissimilar data outside the ball. Currently, the existing deep hashing methods based on hash table lookups pay too much attention to similar samples outside the ball and ignore the learning of dissimilar samples inside the ball, leading to a biased model. In this paper, we introduce discriminatory penalty into the exponential loss functions to optimize the Hamming space, leading to Exponential Hashing with Discriminatory Penalty (EHDP), which discriminately penalizes similar/dissimilar data inside and outside the Hamming ball. Technically, EHDP capitalizes on exponential function to discriminatively encourage similar/dissimilar data approaching/away and to up-weight/down-weight the dissimilar data inside/outside the ball. Extensive experiments demonstrate that the proposed EHDP obtains superior results on three benchmark datasets.