Deep hashing with mutual information: A comprehensive strategy for image retrieval

While deep learning has significantly enhanced image retrieval, several challenges persist. These include inherent information loss during the hash-coding layer, the adverse impact of hard samples during training, and the complexity of generating hash centers within the real number field. To address these issues, we introduce a novel approach termed Deep Hashing with Mutual Information (DHMI), leveraging principles of mutual information. DHMI employs mutual information theory to entire deep hashing to address the aforementioned challenges, respectively: 1) Multi-stage hash-coding layer construction: DHMI constructs the hash-coding layer as multi-stage process (high-dimensional real space-* high-dimensional Hamming space-* low-dimensional Hamming space) that preserves the mutual information throughout the spatial transformation, thereby reducing the information loss during hash-coding. 2) Resilient training process: DHMI constrains the mutual information between the output hash code and the pre-generated hash centers, and develops a loss function capable of automatically adjusting the weight of hard samples, making the training process more resilient. 3) Efficient hash center generation: DHMI utilizes his own hash-coding layer as the center generation network. By inputting a Hadamard matrix instead of image feature, DHMI avoids local optima associated with various discrete optimization algorithms and generates linearly independent hash centers. Leveraging mutual information, DHMI surpasses state-of-the-art (SOTA) deep hashing methods, demonstrating exceptional retrieval performance.