基于Faster R-CNN的排水管道缺陷检测研究

为了克服传统深度学习在排水管道缺陷检测方面识别正确率较低的缺点，在Faster R-CNN算法基础上，利用聚类分析方法改进候选区域设置，提出一种优化的排水管道缺陷检测模型，并采用VGG、AlexNet、GoogleNet、ResNet代替Faster R-CNN网络中的特征提取层进行模拟计算。计算结果表明，K-means方法的最优类别数为5，虽然ResNet网络训练时间成倍增加，但其识别正确率达到0.89，比VGG网络提高了0.14。优化后的Faster R-CNN网络有效提高了排水管道缺陷检测的识别正确率。

Research on Defect Detection of Drainage Pipeline Based on Faster R-CNN

In order to overcome the disadvantage of low recognition accuracy rate of traditional deep learning in defect detection of drainage pipelines， an optimized defect detection model of drainage pipelines based on Faster R-CNN algorithm is proposed， which can improve the setting of anchor box with cluster analysis. VGG， AlexNet， Google Net and ResNet are used to replace the feature extraction layer in Faster R-CNN network， then the simulation calculation is carried out. The results show that the optimal cluster number of K-means method is 5， and while time consuming， the ResNet network has the recognition accuracy rate of 0.89， which is 0.14 higher than VGG network. The optimized Faster R-CNN network effectively improves the recognition accuracy of drainage pipeline defect detection.