Attention Multi-scale Fusion Hashing for Image Retrieval

Abstract

With the rapid development of the Internet and the dramatic increase of image data, image retrieval become a research hot topic. Deep hashing is a binary retrieval method for large-scale data, which combines the powerful feature learning capabilities of neural networks and the inherent advantages of small storage and fast retrieval of hash codes and has become a powerful tool for solving multimedia retrieval problems in recent years. Image feature extraction is the basis of image retrieval and has an important impact on the results of image retrieval. To obtain better image features, this paper proposed an attention multi-scale fusion module to enhance the global and local features and introduced SoftPool to retain more critical features. This algorithm can better solve the problem of gradient disappearance through the SoftSign function. The experiments show that this algorithm performs well in hashing retrieval, which verifies the effectiveness of the algorithm.