An Improved Deep Supervised Hashing Method for Hamming Space Retrieval

Abstract

Due to its storage and computation efficiency, hashing has attracted extensive research on large-scale image retrieval tasks in recent years. This work focuses on Hamming space retrieval which enables the most efficient constant-time search by hash table lookups. In this paper, a novel deep supervised hashing method is proposed to generate highly concentrated hash codes based on a redesigned cross-entropy loss function. We also employ a regularizer term to mitigate the discrepancy between the Euclidean distance and the Hamming distance. Extensive experimental results demonstrate the superior performance of our method compared with existing hashing methods on two large-scale image datasets.