基于Deeplabv3+的排水管道缺陷检测与语义分割

Drainage Pipeline Defects Detection and Semantic Segmentation Based on Deeplabv3+

传统计算机视觉技术应用于排水管道缺陷检测和评估,存在识别类型单一、时效性差、判断准确率低等问题,无法满足现代排水管道多缺陷共存、实时性强和精准检测的需求。近年来兴起的深度学习神经网络技术,具有强大的数据特征学习和处理能力。为此,提出了基于Deeplabv3+卷积神经网络的管道缺陷检测及语义分割方法,实现对排水管道缺陷的多类型检测、空间定位和几何属性分割。分别比较了ResNet-18、ResNet-50、Mobilenet_v2、Xception和InceptionResnet_v2这5类骨架特征提取网络对缺陷检测和语义分割的影响作用。结果表明,ResNet-50的识别分割性能优于其他网络,准确率达到89.8%,平均交并比和加权交并比分别为53.2%和83.9%,分割速率为12.50帧/s。这为排水管道缺陷的智能检测与分割提供了新的技术支撑和手段。

The application of traditional computer vision technology in inspection and evaluation of CCTV sewer defects has many problems,such as incapable of multiple defects identification,poor detection effects and low detection accuracy,which cannot solve the tasks on multiple defects,real-time and accurate detection of drainage defects.In recent years,the emerging deep learning neural network technology has powerful data feature learning and processing capabilities.This paper proposes a pipeline defect detection and semantic segmentation method based on Deeplabv3+convolutional neural network,which is used for multi-type detection,locating and geometric attribute segmentation of sewer defects.The impacts and mechanism of five types of backbone feature extraction networks,including ResNet-18,ResNet-50,Mobilenet_v2,Xception and InceptionResnet_v2,on defect detection and segmentation were analyzed respectively.The experimental results showed that the recognition and segmentation performance of ResNet-50 outperformed other networks.The accuracy rate was 89.8%,the mean and weighted mean intersection over union are 53.2%and 83.9%,respectively,and the segmentation rate is 12.50 frames/s.This method proposed can provide technical support and means for intelligent detection and segmentation of sewer defects.