服装成衣尺寸AI辅助检测

AI-Assisted Detection of Garment Dimension

摘要: 针对传统服装成衣尺寸人工检测效率低、精度易受人为因素影响、难以适配规模化生产的问题，本研究提出基于YOLO改进的算法SV-YOLO，完成成衣尺寸测量任务。结合YOLOv8设计双主干网络结构以增强特征点提取能力，提出SEMA注意力机制利用残差学习思想捕获复杂特征信息；将SV策略融入YOLOv8模型后，关键点定位任务的PointTop1指标提升7.2%、PointTop2指标提升6.8%；在此基础上优化得到的SV-YOLOv10模型，PointTop1指标提升3.1%、PointTop2指标提升3.0%，SV-YOLOv11模型的PointTop1与PointTop2指标均提升3.1%。对比SE、CBAM、CA、EMA等主流注意力机制的效果，SEMA注意力机制表现最优，其使PointTop1指标提升3.1%、PointTop2指标提升3.8%。

Abstract: Addressing the issues of low efficiency, susceptibility to human factors affecting accuracy, and difficulty in adapting to large-scale production in traditional manual detection of garment dimensions, this study proposes an improved algorithm based on YOLO, named SV-YOLO, to complete the task of garment dimension measurement. By combining YOLOv8, a dual-backbone network structure is designed to enhance the ability to extract feature points, and the SEMA attention mechanism is proposed to capture complex feature information using the idea of residual learning. After integrating the SV strategy into the YOLOv8 model, the PointTop1 metric for keypoint localization task improved by 7.2%, and the PointTop2 metric improved by 6.8%. Based on this, the optimized SV-YOLOv10 model achieved a 3.1% improvement in the PointTop1 metric and a 3.0% improvement in the PointTop2 metric, while the SV-YOLOv11 model achieved a 3.1% improvement in both the PointTop1 and PointTop2 metrics. Compared with mainstream attention mechanisms such as SE, CBAM, CA, and EMA, the SEMA attention mechanism performed the best, improving the PointTop1 metric by 3.1% and the PointTop2 metric by 3.8%.