LAUSR

In recent years, deep learning of hash codes for fast image retrieval have achieved excellent performance. Although the off-the-shelf methods achieve promising performance on images of good quality, their performances may degrade greatly on image of low-quality and low-resolution. In this paper, we propose a novel end-to-end deep cooperative enhancement hashing network (DCEN) for low-resolution image retrieval. It aims to promote semantic information of low-resolution images with super-resolution techniques, so as to achieve similar semantic features as high-resolution images. The proposed framework mainly consists of two main components: an image semantic enhancement network and an image hashing network. Specifically, the semantic enhancement network is proposed to generate super-resolved images from low-resolution images, which improves hashing performance of low-resolution images. And the hashing network is presented to not only assist the training of semantic enhancement network, but also to represent images as hash codes. Finally, we adopt an alternative training method for these two networks to greatly reduce their coupling degree, so that the hashing network can still maintain promising performance on high-resolution images. In addition, we propose a bridging strategy to add more semantic information for the super-resolution image. Extensive experiments show that our method achieve state-of-the-art performance on low-resolution images retrieval and still maintains the excellent hashing representation ability for high-resolution images.