Attention-Aware Invertible Hashing Network

In large-scale image retrieval tasks, hashing methods based on deep convolutional neural networks (CNNs) play an important role due to elaborate semantic feature representation. However, they usually progressively discard information during feature transformation, thus leading to incomplete and unsatisfactory hashing codes for image retrieval. This study tries to design an invertible architecture to maintain image information, meanwhile focus on necessary parts of image features. Consequently, in this paper, we propose a novel attention-aware invertible hashing network (AIHN) for image retrieval. By invertible feature representations, the final hash codes can be completely obtained from input images without any information loss. For highlighting informative regions, we present a novel attention-aware invertible block as the basic module of AIHN, which can promote generalization ability by spatial attention mechanism. Extensive experiments conducted on benchmark datasets demonstrate the effectiveness of our invertible feature representation on hash code generation, and show the promising performance on image retrieval of our methods against the state-of-the-arts.