Multi-branch and multi-loss learning for fine-grained image retrieval

To effectively address the problem of low accuracy of fine-grained image retrieval due to significant intra-class differences and small inter-class differences, we propose a novel and highly reliable fine-grained deep hashing learning framework dubbed MBLNet to accurately retrieve fine-grained images. Specifically, we propose (i) a dual-selected significant region erasure method for generating compact binary codes for fine-grained images; (ii) a dual filtering object location method for mining discriminative local features; and (iii) a new multi-stage loss function for optimizing network training. We conducted extensive experiments on three fine-grained datasets, Stanford Cars, FGVC-Aircraft, and CUB-200-2011, and achieved mAP results of 89.3%, 87.2%, and 80.6%, respectively. Additionally, the ablation study demonstrates that both the dual-selected significant region erasure method and the dual filtering object location method contribute to the improved accuracy of fine-grained image retrieval, further validating the effectiveness of the proposed method. Code can be found at https://github.com/luhongchun/MBLNet.git.