基于YOLO v11的轻量化安全帽佩戴检测算法

A Lightweight Safety Helmet Wearing Detection Algorithm Based on YOLO v11

摘要: 在建筑、采矿、勘探等工地施工场景中，安全帽佩戴检测算法是预防工人生产管理安全的重要手段，但现有目标检测模型存在检测精度低、参数量多等问题。本研究提出了一种基于YOLO v11框架LH-YOLO模型，旨在实现精度与效率的高度平衡。首先使用StarNet作为骨干网络，大幅降低计算复杂度；其次，设计了C3k2_SN模块特征融合颈部，模块采用星形跨阶段部分卷积，高效利用参数的同时，增强了特征表达的能力；最后构建了LSCD轻量化共享卷积头，并引入PIoUv2损失函数对边界框定位进行精细优化，进一步压缩检测头参数并优化预测。实验结果表明，在公开数据集SHWD (Safety Helmet Wearing Dataset)上进行实验，改进模型精确率达到93.4%，与原模型参数量降低了32.8%、十亿次的浮点计算量(GFLOPs)下降了34.8%、平均精度提升1.4%，兼顾高检测性能和实时性要求，更适用于实际生产环境的部署与应用。

Abstract: In construction, mining, exploration, and other construction site scenarios, the safety helmet wearing detection algorithm is an important means to prevent safety incidents in worker production management. However, existing object detection models suffer from issues such as low detection accuracy and a large number of parameters. This study proposes an LH-YOLO model based on the YOLO v11 framework, aiming to achieve a high balance between accuracy and efficiency. Firstly, StarNet is adopted as the backbone network, which significantly reduces computational complexity. Second, a feature fusion neck with the C3k2_SN module is designed. The module adopts star-shaped cross-stage partial convolution, which efficiently utilizes parameters while enhancing the capability of feature expression. Finally, an LSCD lightweight shared convolutional head is constructed, and the PIoUv2 loss function is introduced to finely optimize bounding box localization, further compressing the parameters of the detection head and optimizing predictions. Experimental results show that when tested on the public dataset SHWD (Safety Helmet Wearing Dataset), the precision of the improved model reaches 93.4%. Compared with the original model, its parameter count is reduced by 32.8%, Giga Floating-Point Operations (GFLOPs) decrease by 34.8%, and average precision (AP) increases by 1.4%. The model balances high detection performance and real-time requirements, making it more suitable for deployment and application in actual production environments.