A lightweight fashion style recognition model based on large kernel separable attention mechanism
Xiaohui Wu, Yanyan Liao, Xin Luo, Juntong Chen, Yihua He, Shuailing Lu · 2025
Fashion style recognition is challenging due to complex features such as global contours, local textures, and fine-grained stitching. To address this, we propose Efficient-Fashion, a lightweight model based on a large kernel separable attention mechanism. We replace YOLO11’s standard convolution with depthwise separable convolution, reducing parameters. The C3k2_MSCB module enhances feature extraction using multi-scale depthwise convolutions, and the C2PSA_LSKA attention mechanism improves global feature capture while reducing parameters. Experiments on the FashionStyle14 dataset show that Efficient-Fashion achieves 78.01% accuracy, surpassing YOLO11s by 1.89%, with 39.1% fewer parameters and 48.8% lower computation cost. The model processes frames 16% faster, at 27.8ms. This work presents an effective solution for balancing accuracy and efficiency in fashion recognition tasks.