基于改进YOLOX-m的安全帽佩戴检测  

Safety Helmet Wearing Detection Based on Improved YOLOX-m

作　　者：王晓龙 江波 WANG Xiaolong;JIANG Bo(Industry Digital Intelligence Division,ECCOM Network System Co.,Ltd.,Shanghai 200127,China;The 32nd Research Institute of China Electronics Technology Group Corporation,Shanghai 201808,China)

机构地区：[1]上海华讯网络系统有限公司行业数智事业部,上海200127 [2]中国电子科技集团公司第三十二研究所,上海201808

出　　处：《计算机工程》2023年第12期252-261,共10页Computer Engineering

摘　　要：安全帽佩戴检测是安全监控系统中的重要组成部分,其检测精度取决于目标分类、小目标检测、域迁移差异等因素。针对现有基于YOLOX-m模型的安全帽佩戴检测算法通常存在分类精度较低、检测目标不完整、轻量化模型性能下降等问题,构建一种基于多阶段网络训练策略的改进YOLOX-m模型。首先对YOLOX-m主干特征网络卷积块的堆叠次数进行重新设计,在减小网络规模的同时最大化模型性能,然后将残差化重参视觉几何组与快速空间金字塔池化相结合,提高检测精度和推理速度。设计一种多阶段网络训练策略,将训练集和测试集拆分成多个组,并结合推理阶段生成的伪标签进行多次网络训练,以减少域迁移差异,获得更高的检测精度。实验结果表明,与YOLOX-m模型相比,改进YOLOX-m模型的推理延迟降低了5 ms,模型大小减少了4.7 MB,检测精度提高了1.26个百分点。The safety helmet wearing detection is a crucial part of the security monitoring system.Its precision depends on object classification,small-object detection,domain transfer discrepancy,and other factors.Existing algorithms based on YOLOX-m for safety helmet wearing detection have drawbacks of reduced classification precision,incomplete detection targets,and degraded performance of lightweight models.An improved YOLOX-m model based on a multi-stage network training strategy is proposed to solve these problems.First,the number of stacks of convolution blocks of the YOLOX-m backbone feature network is redesigned to maximize the performance of the model while reducing the network.Next,the Residual Re-parameterized Visual Geometry Group(Res-RepVGG)is combined with Spatial Pyramid Pooling-Fast(SPPF)to improve the detection accuracy and reasoning speed.In addition,a multi-stage network training strategy is proposed,which divides the training and test sets into multiple groups and combines the pseudo labels generated in the inference stage for multiple network training to reduce the domain transfer difference and improve the detection accuracy.The experimental results show that compared with YOLOX-m,the improved YOLOX-m exhibits improved performance in helmet wearing detection in three aspects:the delay is reduced by 5 ms,the model size is reduced by 4.7 MB,and the average accuracy is improved by 1.26 percentage points.

关 键 词：安全帽佩戴检测 深度学习 残差化重参视觉几何组 快速空间金字塔池化 多阶段网络训练策略 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]