基于yolov5s的改进安全帽检测算法  

Improved helmet detection algorithm based on yolov5s

作　　者：姚庆安[1] 宋铭轩 冯云丛 乔石丽 张语然 YAO Qingan;SONG Mingxuan;FENG Yuncong;QIAO Shili;ZHANG Yuran(School of Computer Science&Engineering,Changchun University of Technology,Changchun 130102,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130102

出　　处：《长春工业大学学报》2024年第2期138-146,共9页Journal of Changchun University of Technology

基　　金：吉林省自然科学基金(YDZJ202201ZYTS422);吉林省科技厅青年成长科技计划项目(20210508039RQ)。

摘　　要：针对现有安全帽检测算法对于远距离目标以及背景复杂的工地场景下的安全帽识别检测精度较低的问题,对yolov5s结构进行改进,主干网络中引入CoorAtt注意力机制增强特征提取能力,加强对重要的小目标信息的关注;然后将原模型中的SPP模块替换成ASPP模块,通过使用空洞卷积层来代替池化层,降低了最大池化导致的特征信息丢失,同时采用不同的扩张率增大感受野,并且有效地提取不同尺度的特征;其次在颈部网络使用BiFPN结构,更高效地对特征信息进行融合;最后通过更改损失函数为WIOU通过引入动态非单调聚焦机制,平衡模型对各质量样本的关注,提高网络的整体性能,从而提高目标检测精度。为了测试算法的有效性,文中在公共数据集Safety Helmet Detection上进行实验。实验结果表明,改进后的yolov5s算法,目标检测mAP达到了88.5%,比改进之前的yolov5s算法提升了2.1%。Aiming at the problem that the existing safety hat detection algorithm has low detection accuracy for the identification and detection of the safety hat in remote targets and construction site scenes with complex background,this paper improved the yolov5s structure,introduced CoorAtt attention mechanism into the backbone network to enhance the feature extraction capability and strengthen the attention on important small target information,and then the SPP module of the original model is replaced by the ASPP module,by using void convolution layer instead of pool layer,the loss of feature information caused by maximum pool is reduced,and the receptive field is enlarged with different expansion rates,and features of different scales are extracted effectively Secondly,the BIFPN structure is used in the neck network to fuse the feature information more efficiently,and finally,the loss function is changed to WIOU by introducing a dynamic non-monotonic focusing mechanism,the balance model focuses on each quality sample to improve the overall performance of the network,thus improving the accuracy of target detection.In order to test the effectiveness of the algorithm,this paper carries on the experiment on the public data set Safety Helmet Detection.Experimental results show that the improved yolov5s algorithm achieves 88.5%mAP detection,which is 2.1%higher than the yolov5s algorithm before the improvement.

关 键 词：目标检测 安全帽检测 YOLO算法 ASPP 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]