基于轻量化模型的钢轨扣件缺陷检测系统

Rail Fastener Defect Detection System Based on Lightweight Model

铁路巡检工作中,嵌入式设备受算力和存储空间的限制,存在使用YOLO V5模型检测钢轨扣件缺陷速度慢、精度较低的问题。通过替换YOLO V5主干卷积网络为MobileNet V3,将网络中的激活函数修改为Mish并融合协同注意力机制,实现模型的轻量化改进。将改进后的模型部署到嵌入式设备Jetson TX2上,使用板载CSI摄像头扫描、拍摄钢轨扣件,并搭载显示屏等设备构成钢轨扣件缺陷检测系统。运行系统,单张扣件图片的检测速度达56.8 ms,准确度在90%以上,并且模型大小仅有9.8 MB,符合占用存储少、检测效果佳的轻量化要求。

During railway inspection work,since embedded devices are limited by computing power and storage space,there are problems of slow speed and low accuracy in detecting rail fastener defects when using YOLO V5.By replacing the backbone convolutional network of YOLO V5 with MobileNet V3,the activation function in the network was modified to Mish and the coordinate attention mechanism was integrated to realize the lightweight improvement of the model.The improved model was deployed on the embedded device Jetson TX2,scanned and captured the rail fastener by an on-board CSI camera,and mounted with devices such as a display screen to form a rail fastener defect detection system.Launching the system shows that the detection speed of a single fastener picture is 56.8 ms,the accuracy is above 90%,and the size of the model is only 9.8 MB,which meets the lightweight requirements of low storage occupancy and good detection effect.