LDSD: A lightweight detection model for fine-grained recognition of dress style

Efficient and accurate fine-grained recognition of dress styles serves as a critical technical foundation for the intelligent operation of the apparel e-commerce industry. However, existing recognition models often struggle to balance detection accuracy with deployment efficiency on resource-constrained edge devices. To address this issue, this study proposes a lightweight detection model, namely Lightweight Dress Style Detection (LDSD). A fine-grained dress dataset covering six core design attributes, including overall silhouette, waistline position, skirt length, collar shape, sleeve length, and sleeve shape, was first constructed with 22 categories in total. Based on the YOLO11n framework, the proposed model was optimized from four aspects: feature extraction, convolution operators, feature fusion network, and detection head design. Specifically, the C3k2-PC module was developed to reduce channel redundancy while preserving fine-grained structural features. An ultra-lightweight convolution named U_GSConv is proposed, which compresses spatial computational overhead via adaptive channel splitting. In addition, a Lite-BiFPN asymmetric bidirectional feature fusion network was developed to enhance multi-scale feature representation, while the LSCD_DSConv detection head combined depthwise separable convolution with a weight-sharing mechanism to improve parameter utilization efficiency. Experimental results demonstrate that the LDSD model reduces its parameter count and floating-point operations (FLOPs) to 1.14 M and 2.7 G, respectively. With a model size of merely 2.5 MB, it achieves a 51.9% reduction relative to the baseline. Meanwhile, LDSD achieves 94.0% mAP@50 and 115.22 FPS on the self-constructed dress dataset, demonstrating a favorable balance between recognition accuracy and inference efficiency. The proposed method provides an efficient and low-cost solution for intelligent apparel sorting, automatic product classification, and personalized recommendation systems in e-commerce scenarios.