Supervised deep quantization for efficient image search

Abstract

Due to the efficiency of compact binary codes in approximate nearest neighbor search for large-scale image retrieval, hashing techniques have received increasing attentions. For most existing hash methods, the suboptimal binary codes are generated, as the hand-crafted feature representation is not optimally compatible with the binary codes. In this paper, we propose a one-stage supervised hashing method to learn high-quality binary codes. We implement a deep Convolutional Neural Network and enforce the learned codes to meet the following criterions: (a) similar images should be encoded into similar binary codes, and vice versa; (b) the binary codes should be evenly distributed; (c) the loss of quantization should be minimized. Experimental comparisons between our method and state-of-the-art algorithms are conducted on CIFAR-10 and NUS-WIDE datasets, and the MAP of our method reaches to 87.67% and 77.48% with 48-bit respectively. It shows that our method can obviously improve the search accuracy.