Deep Supervised Hashing with Information Loss

Recently, deep neural networks based hashing methods have greatly improved the image retrieval performance by simultaneously learning feature representations and binary hash functions. Most deep hashing methods utilize supervision information from semantic labels to preserve the distance similarity within local structures, however, the global distribution is ignored. We propose a novel deep supervised hashing method which aims to minimize the information loss during low-dimensional embedding process. More specifically, we use Kullback-Leibler divergences to constrain the compact codes having a similar distribution with the original images. Experimental results have shown that our method outperforms current stat-of-the-art methods on benchmark datasets.