Fashion-YOLO: clothing style recognition based on attention mechanism and programmable gradient

Clothing style recognition is challenging due to large variations in global appearance, local design, and fine-grained textures, and existing studies often rely on relatively homogeneous datasets that limit generalization in practical applications. To address these issues, we introduce ShanghaiFashionStyle17, an expert-curated clothing style dataset designed to reflect modern fashion diversity, comprising 17 style categories and 33,184 images. We further propose Fashion-YOLO, an enhanced YOLOv8-based model for clothing style recognition. Specifically, Fashion-YOLO integrates the SENetV2 channel attention mechanism to strengthen global and local detail feature extraction, employs Programmable Gradient Information (PGI) to optimize gradient flow and improve the learning of fine-grained style cues, and incorporates depthwise separable convolutions (DWConv) to reduce parameter count while maintaining effectiveness. Experiments on ShanghaiFashionStyle17 and FashionStyle14 show that Fashion-YOLO achieves 81.7% and 78.7% Top-1 accuracy, respectively, outperforming the baseline by 4.5% and 3.8%. In addition, we report practical runtime indicators under a unified setting, conduct error analysis using a normalized confusion matrix and representative failure cases, and perform cross-dataset transfer evaluation on semantically aligned overlapping categories between the two datasets to assess generalization. Overall, Fashion-YOLO demonstrates a favourable accuracy–efficiency trade-off and ShanghaiFashionStyle17 supports learning transferable style representations.