Bit-wise attention deep complementary supervised hashing for image retrieval

Deep hashing is effective and efficient for large-scale image retrieval. Most of existing deep hashing methods train a single hash table by utilizing the output of the penultimate fully-connected layer of a convolutional neural network as the deep feature of images. They concentrate on the semantic information but neglect the fine-grain image structure. To address this issue, this paper proposes an advanced image hashing method, Bit-wise Attention Deep Complementary Supervised Hashing (BADCSH). It is an end-to-end system that trains a sequence of hash tables in a boosting manner, each of which is trained by correcting errors caused by all previous ones. Features from different levels of the network are used to train different hash tables. The hash table trained with features at one level reveals a level of semantic content of the image, while the hash table trained with features at a lower level contains structural information of the image that makes up the semantic content. Moreover, the hash layer is used as an embedded layer of the network to generate hash codes. A dense attention layer is added to the hash layer to treat various hash bits differently, in order to reduce hash code redundancy and maximize overall similarity preservation. Finally, the hash tables trained on different levels of features are fused by weights computed based on their respective performance. Experiments on three real-world image databases demonstrate that the proposed method achieves the best performance among state-of-the-art comparative hashing methods.