基于高效通道注意力模块(ECA)和YOLOv5的图像检测方法研究

An Image Detection Method Based on ECA and YOLOv5

佩戴安全帽是人们在施工建设中的一项重要保护措施,它可以有效保障人们的生命财产安全。安全帽检测系统也已经成为了很多施工场所的必备的基础设施,为了改善YOLOv5不能通过权重进行聚焦,从而生成有明显辨识度的特征,进而影响安全帽检测准确度的问题,我们在YOLOv5中引入了注意力模块,保证了卷积过程中的特征提取,并且使得图像得到优化,提升了安全帽检测结果的准确性和模型性能。并且我们对比了原YOLOv5、添加了ECA(Efficient Channel Attention)高效通道注意力模块、添加了SEA(Squeeze-andExcitation attention)注意力模块和添加了压缩激励SEL(Squeeze and Excitation Layer)注意力模块的精确率P/%、召回率R/%、mAP@0.5和mAP@0.5:0.95,实验结果表明添加了ECA(Efficient Channel Attention)通道注意力模块的ECA-Yolov5算法相较于原YOLOv5算法P/%、R/%、mAP@0.5、mAP@0.5:0.95分别提升了0.5、0.6、0.5、0.2。由此结果表明引入高效通道注意力模块(ECA)的YOLOv5安全帽识别算法更有能力进行安全施工的检测,提升了施工的安全性。

Wearing safety helmet is an important protection measure for people in construction,it can effectively protect people's life and property safety.The hard hat detection system has also become a necessary infrastructure for many construction sites.In order to improve the problem that YOLOv5 cannot focus by weight,so as to generate features with obvious recognition,thus affecting the accuracy of hard hat detection,we introduce an attention module into YOLOv5 to ensure feature extraction in the convolution process.Improve the accuracy of test results and model performance of safety helmet.We also compare the original YOLOv5.We add ECA(Efficient Channel attention),SEA(Squeeze and Excitation)attention module.The compression excitation SEL(Squeeze and Excitation Layer)Precision P/%,recall R/%,mAP@0.5 and mAP@0.5 of the attention module:0.95,The experimental results show that the efficiency Channel Attention module added by ECA-Yolov5 algorithm is compared with the original YOLOv5 algorithm P/%,R/%,mAP@0.5,mAP@0.5:0.95 Increases of 0.5,0.6,0.5 and 0.2,respectively.The results show that the YOLOv5 safety hat recognition algorithm introduced with high efficiency channel attention module(ECA)is more capable of safety construction detection.Improve the safety of construction.