Abstract

Hashing, which refers to the binary embedding of high-dimensional data, has been an effective solution for fast nearest neighbor retrieval in large-scale databases due to its computational and storage efficiency. Recently, deep learning to hash has been attracting increasing attention since it has shown great potential in improving retrieval quality by leveraging the strengths of deep neural networks. In this paper, we consider the problem of supervised hashing and propose an effective model (i.e., DHA), which is able to generate compact and discriminative binary codes while preserving semantic similarities of original data with an adaptive loss function. The key idea is that we scale and shift the loss function to avoid the saturation of gradients during training, and simultaneously adjust the loss to adapt to different levels of similarities of data. We evaluate the proposed DHA on three widely-used benchmarks, i.e., NUS-WIDE, CIFAR-10, and MS COCO. The state-of-the-art image retrieval performance clearly shows the effectiveness of our method in learning discriminative hash codes for nearest neighbor retrieval.