Deep hashing via multi-scale and multi-directional pooling for image retrieval

Deep Hashing methods have been widely used for large-scale image retrieval due to its advantages in retrieval efficiency and accuracy. Recent methods cannot effectively capture the scale variation and complex distribution of image features in the feature extraction process. As a result, these methods were easily affected by the environment and the retrieval accuracy is not high. In this regard, we propose multi-scale and multi-direction pooling for deep hashing (MMDH). Specifically, the proposed method uses a multi-direction strip pooling module (MSPM) so that it can also fuse the features in the diagonal direction. Also, we combine MSPM with pyramid pooling which capture multi-scale features to build a hybrid pooling module. The input image is subjected to the hybrid pooling operation to extract multi-scale and multi-direction image features which can help the network improve its feature extraction ability. The atrous spatial pyramid pooling operation is used to retain the multi-scale features further to help the model analyze the image structure flexibly and effectively. The experimental results show that the proposed method can perform well on two public datasets. In addition, to verify the generality of the results, we also conducted experiments on the cloth dataset Fabric. The results prove that the proposed method can better extract multi-scale and multi-direction image feature information in various situations.