基于改进YOLOv5s的齿轮表面缺陷检测

Gear surface defect detection based on improved YOLOv5s

为解决传统齿轮表面缺陷检测精度低、速度慢等问题,设计了一种基于通道和空间的注意力机制的齿轮表面缺陷检测方法。在YOLOv5s网络模型的基础上引入卷积注意力模块,对特征在通道维度和空间维度上进行融合增强,增强缺陷区域特征并抑制背景区域特征,提高小目标的检测精度;同时改进了非极大值抑制的后处理方法,改进后的方法(DIOU_NMS)将预测框与真实框的重叠区域和2个框之间的中心点距离作为抑制原则,提升复杂背景下目标的检测精度。实验结果表明,该方法的平均精度均值mAP_0.5为90.3%,相比YOLOv5s提升了1%,检测速度FPS为75 f/s,模型大小为14.8 MB,满足齿轮表面缺陷检测实时性和准确性的需求。

To solve the problems of low precision and slow speed of traditional gear surface defect detection,a gear surface defect detection method based on channel and spatial attention mechanism was designed.On the basis of the YOLOv5 s network model,the convolutional attention module was introduced,and the features were fused and enhanced in the channel and space dimension,which enhanced the characteristics of the defect area and suppressed the characteristics of the background area,and improved the detection accuracy of small targets.At the same time,DIOU_NMS was used as the post-processing method,and the overlapping area between the prediction frame and the real frame and the center point distance between the two frames were taken as the suppression principle to improve the detection accuracy of the target in complex background.The experimental results show that the mAP_0.5 of the proposed method is 90.3%,1%higher than that of YOLOv5 s,the FPS is 75 f/s,and the model size is 14.8 MB,which can meet the requirements of real-time and accuracy of gear surface defect detection.