MR-GA:一种基于实例分割的地下排水管道缺陷评估方法

MR-GA:An instance segmentation-based method for assessing underground drainage pipe defects

目前地下排水管道检测后的缺陷评估主要采用人工方式进行,不仅费时、费力,智能化程度低,而且容易受到人工经验与视觉疲劳的影响,导致遗漏与误判,影响结果的准确性。近年来迅速兴起的实例分割技术,具有强大的数据特征学习和描述能力。提出了一种基于Mask R-CNN实例分割网络的管道缺陷评估MR-GA(Mask R-CNN-Grading Assessment)方法。首先,自建样本数据集,对Mask R-CNN进行训练构建模型;其次,利用构建好的模型对输入的管道检测帧进行缺陷分类和实例分割;在此基础上,结合《城镇排水管道检测与评估技术规程》,根据各类管道缺陷特征,制定出缺陷量化分级评估方案;最后,按此方案进行参数计算及分级评估。将MR-GA方法应用于实际工程项目,与人工评估结果相比,缺陷类别识别准确率达到91.34%,缺陷分级准确率达到88.75%。

At present,the evaluation of defects after underground drainage pipe inspection is mainly carried out manually,which is not only time-consuming and laborious,but also low in intelligence,and easily affected by manual experience and visual fatigue,leading to omissions and misjudgments and affecting the accuracy of results.In recent years,instance segmentation technology has emerged rapidly,with powerful data feature learning and description capabilities.In this paper,we propose a MR-GA(Mask R-CNN-Grading Assessment)method for pipeline defect assessment based on Mask R-CNN instance segmentation network.Firstly,we construct a model by training Mask R-CNN with self-built sample data set.Secondly,we use the constructed model to classify and segment the defects of the input pipe inspection frames.On this basis,we develop a quantitative grading assessment scheme based on the characteristics of various types of pipe defects in conjunction with the Technical regulations for inspection and assessment of urban drainage pipes.Finally,we carry out parameter calculation and grading assessment according to this scheme.Finally,the parameters are calculated and graded according to this scheme.When the MR-GA method was applied to the actual project,the accuracy of defect category identification reached 91.34%and the accuracy of defect grading reached 88.75%compared with the manual evaluation results.