DDSH: Deep Distribution-Separating Hashing for Image Retrieval

摘要

With the rapid growth of web images, binary hashing method has received increasing attention due to the storage efficiency and the ability for fast retrieval. Recently, deep hashing methods have achieved the state-of-the-art performance by utilizing deep neural networks in hash code learning. Most of these methods are trained with the supervision of triplet labels or pairwise relationships. In this paper, we propose a deep hashing framework called deep distribution-separating hashing (DDSH) method. The main novelty of our learning framework lies in the supervision which enforces to separate the distribution of similar pairs from the distribution of dissimilar pairs. In this way, the gap between similar pairs and dissimilar pairs is enlarged. Experimental results show that our proposed deep hashing method outperforms state-of-the-art approaches on two widely used benchmark datasets: CIFAR-10 and PASCAL VOC 2007.