Fine-Grained Clothing Classification

Fine-Grained Clothing Classification (FGCC) seeks to accurately distinguish specific subcategories of apparel. Compared to other Fine-Grained Visual Classification (FGVC) tasks, FGCC is more challenging due to pose variations and clothing style differences that cause significant deformation and occlusion. Moreover, the coupling of multiple attributes (e.g., color, pattern, and texture) not only blurs inter-class boundaries but also amplifies intra-class differences. To address these challenges, this study makes the following contributions: (1) An Adaptive Group Interaction Fusion (AGIF) module is proposed to adaptively group features and enhance representation through inter-group semantic interactions. (2) A Channel Attention with Feature Enhancement (CAFE) mechanism is introduced to refine channel-wise feature importance, enabling the model to focus on subtle but discriminative details while suppressing irrelevant signals. (3) A Spatial Feature Selection (SFS) mechanism is designed to emphasize key regions via multi-scale spatial attention, improving the model\u2019s sensitivity to fine-grained spatial patterns. (4) To support FGCC under real-world conditions, we construct FashionNet, a large-scale fine-grained fashion dataset comprising 892,841 clothing images across 162 subcategories and 8 attribute dimensions (e.g., sleeve type, collar shape, contour, fabric). Compared with existing datasets, FashionNet features higher category density, complex backgrounds, diverse viewpoints, and up-to-date fashion trends. It provides a rich and realistic benchmark for weakly supervised fine-grained classification. Comprehensive experiments conducted on FashionNet and two standard benchmarks demonstrate that our proposed framework achieves state-of-the-art classification performance, with a TOP-1 accuracy of 89.19%. These results highlight the model\u2019s effectiveness and robustness in tackling real-world FGCC challenges.