基于卷积神经网络的排水管道缺陷智能检测与分类

Intelligent Detection and Classification of Drainage Pipe Defects Based on Convolutional Neural Networks

针对传统排水管道缺陷检测中需要消耗大量人力的问题,基于人工智能算法——卷积神经网络(CNN),建立自动检测和评估排水管道缺陷的智能系统。对CCTV视频图像中六种常见的管道状态(裂缝、错口、障碍物、残墙坝根、树根和正常类别)进行模型学习、训练和测试。CNN模型训练和验证的正确率分别为100%和97%,六类管道状态的平均识别准确率达到90%,证明所构建的模型在不需要相关检测专业知识的情况下,可以很好地识别本研究中的管道缺陷类型。其中,CNN模型对树根和错口的检测具有较高置信度,其次是残墙坝根和裂缝,障碍物和正常类别的分类精度最低。深度学习在排水管道缺陷自动检测领域具有可行性,模型具有良好的泛化能力,可为管道缺陷检测提供科学、准确的检测工具。

Traditional drainage pipe defect detection needs a lot of manpower.To cope with the problem,a system for automatic detection and evaluation of the drainage pipeline defects was established based on an artificial intelligence algorithm—convolutional neural networks(CNN).Six common pipeline defects(crack,disjoint,obstacle,residual wall,tree root and normal category)observed by CCTV video images were learned,trained and tested by the model.The training and validation accuracies of the CNN model were 100%and 97%,respectively,and the average recognition accuracy of the six kinds of pipeline defects reached 90%,which proved that the established model could well identify the defect types without the need of relevant detection expertise.The CNN model had a high confidence in the detection of the tree roots and disjoints,followed by the residual walls and cracks,and the classification accuracy of the obstacles and the normal type was the lowest.The deep learning is feasible in the field of automatic detection of the drainage pipe defects,and the model has good generalization ability,which provides a scientific and accurate detection tool for the detection of the pipe defects.