针对YOLO v5l(you only look once version 5 large)算法对于小目标、少样本且背景复杂的排水管道缺陷图像检测的精度低、误检和漏检率较高等问题,提出了一种基于YOLO v5l-Im算法的排水管道缺陷检测改进方法.做了三点改进:首先提出了 Focal-EIoU(focal embedding intersection over union)损失函数,有效提升了检测模型的性能;其次为增强检测模型对小目标缺陷的检测效果,减少缺陷误检和漏检的概率,将骨干网络中浅层特征图融合到双向特征金字塔网络(bidirectional feature pyra-mid network,BiFPN)中,增加针对小目标的预测层;最后在YOLO v5l中引入坐标注意力机制(coordinate attention,CA),提高模型对图像中感兴趣区域的敏感程度,减少冗余背景信息的干扰.3种改进对平均检测准确率(mean average precision,mAP)的提升分别为2.0、2.9、5.9个百分点.将三种有效改进融合到一起,检测结果表明:本文提出的YOLO v5l-Im模型的mAP达到了92.1％,较原模型的85.5％提升了 6.5个百分点.由此可见,所做的改进有效增强了 YOLO v5l对排水管道缺陷的检测能力.
Detection and Effectiveness of Drainage Pipeline Defects Based on YOLO v5l-Im
An enhanced method for the detection of defective drainage pipes based on the YOLO v5l-Im algorithm was proposed.This method aims to address the low accuracy,high false detection,and leakage rate of the YOLO v5l(You Only Look Once version 5 large)algorithm for the detection of defective images of drainage pipes with small targets,few samples,and complex backgrounds.There are three enhancements.Initially,the Focal-EIoU(focal embedding intersection over union)loss function is suggested,which significantly enhances the detection model's performance.Secondly,the shallow feature map in the backbone network is fused into BiF-PN(bidirectional feature pyramid network)to increase the prediction layer for small targets,improving the detection effect of the detec-tion model on small target defects and lowering the likelihood of defect misdetection and omission.Lastly,YOLO v5l introduces CA(co-ordinate attention)mechanism to lessen the interference of redundant background information and enhance the model's sensitivity to the image's region of interest.The value of mAP(mean average precision)showes improvements of 2.0,2.9,and 5.9 percentage points,re-spectively,among the three improvements.The test findings indicate that the YOLO v5l-Im model suggested in this work obtains a mAP value of 92.1％,which is 6.5 percentage points higher than the 85.5％of the original model after fusing the three effective modifica-tions.It is evident that the enhancements significantly increase the YOLO v5l's capacity to identify flaws in drainage pipes.
NETL
NSTL
万方数据
