Abstract:

Construction， mining， exploration and other industries have mandatory regulations on helmet wearing in production. The helmet wearing detection algorithms have been widely used in the above industries， but the existing algorithms have problems such as too many parameters， high complexity and poor real-time performance. Therefore， a lightweight safety helmet wearing detection algorithm YOLO v8-s-LE was proposed on the basis of YOLO v8 （You Only Look Once v8）. Firstly， the LAD （Light Adaptive-weight Downsampling） method was designed，so that compared with the original YOLO v8 algorithm， the proposed algorithm reduced floating-point computation significantly. Then， the efficient multi-scale convolution C2f_EMC （C2f_Efficient Multi-Scale Conv） method was used to extract multi-scale feature information， which increased the depth of the network effectively， made the neural network take into account both shallow and deep semantic information， and further improved the expression ability of the algorithm for feature information. Experimental results show that compared with YOLO v8-s algorithm on the public dataset SHWD （Safety Helmet Wearing Dataset）， the proposed algorithm has the parameters reduced by 77%， the floating-point computation reduced by 73%， and the precision reached 92.6%， verifying that the algorithm takes into account the requirements of accuracy and real-time performance， and is more suitable for the deployment and application in actual production environments.

Key words: YOLO v8 (You Only Look Once v8), safety helmet wearing detection, lightweight, multi-scale convolution, feature fusion

摘要：

建筑、采矿、勘探等行业对生产环节的安全帽佩戴有着强制性规定，安全帽佩戴检测算法在上述行业得到广泛应用，然而现有算法存在参数量大、复杂度高及实时性差等问题。因此，提出一种基于YOLO v8的轻量化安全帽佩戴检测算法——YOLO v8-s-LE。首先设计了轻量化自适应权重下采样（LAD）方法，相较于原始YOLO v8算法，该算法的参数量和浮点运算量显著下降；然后使用高效多尺度卷积C2f_EMC（C2f_Efficient Multi-scale Conv）方法提取多尺度特征信息，从而有效增加了网络深度，使神经网络兼顾了浅层和深层语义信息，并进一步提高了算法对特征信息的表达能力。在公开数据集SHWD （Safety Helmet Wearing Dataset）上与YOLO v8-s算法对比的实验结果表明，所提算法的参数量减少了77%，浮点运算量下降了73%，精确率达到92.6%，兼顾准确性和实时性要求，更适用于实际生产环境的部署与应用。

关键词: YOLO v8, 安全帽佩戴检测, 轻量化, 多尺度卷积, 特征融合