改进YOLOX-S算法的安全帽和口罩检测  

Improved YOLOX-S Algorithm for Helmet and Mask Detection

作　　者：童晓东 李兆飞 TONG Xiaodong;LI Zhaofei(College of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin,Sichuan 644000,China;Key Laboratory of Artificial Intelligence in Sichuan Province,Yibin,Sichuan 644000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644000 [2]人工智能四川省重点实验室,四川宜宾644000

出　　处：《宜宾学院学报》2023年第6期17-24,共8页Journal of Yibin University

基　　金：国家自然科学基金项目(11705122);四川省科学技术厅科技计划项目(2021YFG0055);自贡市重点科技计划项目(2019YYJC15)。

摘　　要：针对建筑工地等危险场景下需要对相关人员佩戴安全帽和口罩进行检测,提出基于改进YOLOX-S算法对安全帽和口罩小目标进行同时、实时检测.首先,在YOLOX-S中CSPLayer结构引入ECA注意力机制,引导模型更加关注小目标信息的通道特征,增强模型对有用特征的利用能力;其次,在主干特征提取网络的三个特征层后添加ConvNext Block模块,增强模型对有用特征的利用能力;最后,在加强特征提取网络中引入BiFPN的加权特征融合机制,将原来concat变为BiFPN_concat,增加了对每个输入特征添加可学习的权值,来学习不同输入特征的重要性,区分特征融合过程中不同特征的重要程度,更好关注待检测的目标信息.实验结果表明,改进后算法的mAP为93.2%,比原始YOLOX-S算法平均精确度提升了3.1%.Aiming at the need to detect construction workers wearing safety helmets and masks in dangerous scenarios such as construction sites,an improved YOLOX-S algorithm was proposed to detect small targets of safety helmets and masks simultane⁃ously and in real time.First,the ECA attention mechanism was introduced into the CSPLayer structure in YOLOX-S,which guided the model to pay more attention to the channel features of small target information and enhanced the model's ability to utilize useful features;secondly,ConvNext Block was added after the three feature layers of the backbone feature extraction net⁃work,and the Block module enhanced the model's ability to utilize useful features;finally,the weighted feature fusion idea of BiFPN was introduced into the enhanced feature extraction network,the original concat was changed to BiFPN_concat,and learnable weights were added to each input feature to learn the importance of different input features,distinguish the importance of different features in the feature fusion process,and pay more attention to the target information to be detected.The experimen⁃tal results show that the mAP of the improved algorithm is 93.2%,which is 3.1%higher than the average accuracy of the original YOLOX-S algorithm.

关 键 词：深度学习 YOLOX-S 安全帽检测 口罩检测 建筑工地 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]