Efficient information utilization for fine-grained hashing

In recent years, fine-grained hashing methods have emerged as a promising solution for fine-grained image retrieval (FGIR) due to their low storage cost and high retrieval efficiency. Researchers have improved model interpretability and retrieval accuracy by establishing correspondence between hash codes and visual attributes. However, these methods still face two key issues: (1) The attention mechanism often leads to the loss of discriminative information during feature extraction, and (2) the category labels of images are not fully exploited. To address these issues, we propose the enhancements–uppression (ES) strategy and the multiclassifier feature learning (ESMCFL). To address the first issue, we introduce the ES strategy, which significantly reduces the loss of discriminative information during feature extraction. To address the second issue, we propose the multiclassifier feature learning (MCFL) method, which improves both label utilization and feature discriminability. We conduct experiments on five fine-grained benchmark datasets (CUB200-2011, Aircraft, NABirds, VegFru, Food101). The experimental results demonstrate the effectiveness of our work.