MDH: Mask-Induced Decoupled Hashing for Fine-Grained Image Retrieval
Abstract
Due to its computational and storage efficiency, image hashing has emerged as a significant research approach for fine-grained image retrieval (FGIR). Since fine-grained images exhibit minimal inter-class variance and significant intra-class diversity, existing image hashing methods typically employ multiple local branches to jointly capture the fine-grained local details of each image. However, as the number of branches increases, the overall computational overhead inevitably rises and noise may also be introduced. Therefore, how to reduce the number of branches while maintaining performance advantages remains a challenging problem. To address these challenges, we propose Mask-induced Decoupled Hashing for FGIR (termed as MDH). Specifically, MDH designs an efficient foreground-background decoupling mechanism that separates foreground and background regions at a coarse level through mask prediction, where foreground features focus on the most discriminative core regions of the category, while background features supplement rich contextual information. Subsequently, MDH introduces a foreground-background weight-sharing Graph Aggregation Module (GAM) to further explore structural relationships and high-order semantic information between regions, modeling spatial dependencies within and between foreground and background regions. Finally, MDH devises a concise and efficient gated mechanism to adaptively fuse foreground and background features, capturing local details with both discriminability and robustness. We conduct extensive experiments on four popular fine-grained image datasets. Experimental results consistently demonstrate that compared with the state-of-the-art image hashing methods, our MDH can achieve significant performance improvement while substantially reducing the model complexity. Taking the large-scale dataset Food101 as an example, MDH delivers a remarkable 5.2% absolute mAP improvement while bringing a 36.0% reduction in parameters and a 46.9% decrease in computational cost (FLOPs) at 12-bit hash code length. The code of this project is available at GitHub: https://github.com/dislie/MDH