Abstract: Hash retrieval has attracted wide attention because of its small storage space and fast retrieval speed. There are two problems in deep hashing methods: Deep hash codes are essentially binary features, and the coding length is short, so their feature representation abilities are limited;In addition, existing deep hashing algorithms cannot directly learn discrete hash codes by backpropagation, and usually relax discrete values to continuous values in their optimization procedure, so this leads to quantization errors. Aiming at above problems, we propose a deep hashing retrieval algorithm combining attention model and bimodal Gaussian distribution. The network structure with spatial and channel attention model, focusing on important features and suppressing unnecessary features, enhances the feature representation abilities of hash codes;To solve the quantization error problem, the bimodal Gaussian distribution with the mean of either +1 or-1 is used as the prior distribution. We refer to the idea of variational auto-encoder, and propose to constrain the hash codes distribution to obey the prior distribution with KL divergence. The mean average precision of our method on three benchmark databases CIFAR-10, ImageNet-100 and NUS-WIDE is better than other methods of comparison, which verifies the effectiveness of the algorithm in this paper.