A deep hashing method based on attention module for image retrieval

Hashing retrieval technique has received widespread attention since its low storage consumption and fast retrieval speed. Image retrieval based on deep hashing method uses image features with high-level semantic information and discriminative expression capabilities. However, there are still some limitations: (1) Compared with the traditional feature vector, the length of the learned hash codes is short, and each bit of the code is -1 or 1, the discrimination capability of feature representation is limited; (2) the existing deep hash algorithms cannot directly learn the discrete hash codes using sign activation function, therefore, they use relaxation-based scheme to learn the hash codes, resulting in a large quantization error. In order to solve the above problems, a deep hashing retrieval algorithm incorporate with attention model is proposed. This algorithm embeds spatial and channel attention models in the feature extraction network, and uses a novel activation function in the hash layer. In addition, a novel triple loss function is further proposed to improve the learning capability of the network. Finally, experimental results on the two benchmark databases, i.e., CIFAR-10 and MINIST datasets, show the effectiveness and superiority compared with the baselines.