MSAFNet: Integrating Multi-Scale and Self-Adaptive Attention for Comprehensive Feature Representation

Fine-grained visual classification (FGVC) is a challenging problem in computer vision, focused on distinguishing visually similar subcategories within broader categories. This paper introduces MSAFNet, an innovative framework that integrates two modules: Multi-Scale Attention with Boost and Suppress (MABS) and Self-Adaptive Feature Integration (SAFI). The MABS module enhances discriminative features while suppressing overly prominent ones, ensuring the network learns from a broader range of complementary regions. The SAFI module further integrates local and non-local features, optimizing feature representation for improved classification. We evaluated our approach using three benchmark datasets: CUB-200-2011, Stanford Cars, and FGVC-Aircraft. Experimental results demonstrate that MSAFNet achieves state-of-the-art performance, outperforming existing methods by efficiently capturing and integrating multi-scale discriminative information. These improvements are attributed to our novel use of attention mechanisms and adaptive feature fusion to enhance fine-grained feature learning.