基于YOLO v5l-Im的排水管道缺陷检测方法及效果分析

针对YOLO v5l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，提出一种基于改进YOLO v5l算法的排水管道缺陷检测方法。做了三点改进：首先提出了基于Focal EIoU的损失函数，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BiFPN特征融合网络中，增加针对小目标的预测层；最后在YOLO v5l中引入CA注意力模块，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。三种改进对平均准确度 mAP 值的提升分别为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：本文提出的改进YOLO v5l模型的mAP值达到了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。

Drainage pipe defect detection and effect based on the improved YOLO v5l

A drainage pipe defects detection approach that is based on the improved YOLO v5l algorithm is presented to solve the problems of low accuracy, high leakage and false detection rates of the YOLO v5l algorithm for drainage pipe defect image detection with small targets, few samples and complex backgrounds. Three improvements are made: firstly, a loss function is suggested based on Focal EIoU, which can effectively raise the capability of detection model; secondly, in order to enhance the detection effectiveness of the model for small objectives and lower the rate of false detection and leakage, the shallow feature map in the backbone network is fused into the BiFPN feature fusion network to add a prediction layer for small objectives; finally, in order to improve the sensitivity of the model to the sensitivity of the model to the interested region in the image and decrease the pleonatic intervention of background information, the CA attention module is introduced in YOLO v5l. The three improvements improve the average accuracy mAP values by 2.0, 2.9, and 5.9 percentage points, respectively. Combining the three effective improvements together, the test results show that the improved YOLO v5l model proposed in this paper achieves an mAP value of 92.1%, which is 6.5 percentage points higher than the 85.5% of the original model. This shows that the improvements made effectively enhance the detection capability of YOLO v5l for drainage pipe defects.