Multi-Scale Fusion and Saliency Suppression Network for Fine-Grained Visual Classification

Fine-grained visual classification (FGVC) relies on accurately capturing subtle differences within similar categories, making it a challenging task in computer vision. Current methods often enhance classification by focusing on salient regions or by fusing multi-level features, but they encounter two main issues: (1) excessive reliance on the most prominent regions while overlooking other informative areas, and (2) limited flexibility in integrating features across different scales. To address these challenges, we propose the Multi-Scale Fusion and Saliency Suppression Network (MSFSN), a framework comprising two core modules that can be seamlessly integrated into convolution neural networks. First, we introduce a Multi-Scale Feature Fusion Module (MSFFM) that performs cross-scale feature integration to better capture contextual semantics and fine details. Second, we implement an Attention Suppression Module (ASM) to downplay the most salient features, encouraging the model to explore other potential discriminative regions. Our method does not require additional annotations, such as bounding boxes or key-points, and is designed for end-to-end training. Experimental evaluations on multiple fine-grained benchmarks demonstrate that our method achieves state-of-the-art performance.