FEDDR:一套实用的地下排水管道缺陷智能检测系统

地下排水管道缺陷检测是地下管线高效管理的基础,也是实现“智慧城市”的关键性问题。针对工程项目中对管道缺陷判别的需要,提出并实现了一套实用的地下排水管道缺陷智能检测FEDDR(frame extracting-detection-duplicate removal)系统,将视频缺陷检测过程分为检测前的视频预处理阶段、缺陷检测模型构建阶段以及缺陷检测优化3个阶段,采用帧间差分算法及VGG16网络对管道视频抽帧处理,筛选出兴趣检测帧,减少待检测数据量;选取YOLOv3为网络主框架,用轻量高效的EfficientNet结构替换原来的主干网络,采用迁移学习策略,用自建数据集Pipe-DATA对其进行训练,建立起高效的管道缺陷检测模型,并在检测帧输出检测结果时采用两次输出的优化策略来防止缺陷漏检;对检测出的缺陷帧图像进行文字识别,去重优化自动生成结果表单。将该方法应用到了某区域的将近3 km的管道视频数据中,共检测出了656个缺陷,与人工判别结果对比,准确率达94.3%,召回率达到98.7%,整个过程一体化完成,大大减少了人工成本,提高了排水管道缺陷的检测效率,具有工程实用性。

FEDDR:A practical intelligent detection system for underground drainage pipeline defects

Underground drainage pipeline defect detection is the basis for the efficient management of underground pipelines, and it is also a vital issue to achieve a "smart city".In order to meet the needs of pipeline defect detection in engineering projects, a practical intelligent detection system FEDDR( Frame Extracting-Detection-Duplicate Removal) was proposed and implemented for underground drainage pipeline defects. It divides the process of video defect detection into three stages: video pre-processing, building a defect detection model, and optimizing defect detection. Firstly, the pipeline video frame extraction is processed by the Inter Frame Difference method and VGG16 network to screen out the candidate detection frames and reduce the amount of data to be detected. Secondly, YOLOv3 network is chosen as the mainframe of deep learning, and its backbone network is replaced with EfficientNet. The transfer learning strategy is used, the self-built data-set Pipe-DATA is used to train it, and an efficient pipeline defect detection model is established. To avoid defects miss-detected, the defect detection results are output two times. Finally, the redundant frames are removed by recognizing characters in the detected defect frames, and the defect list is obtained. This method is applied to the video data of about 3 km of pipes in one region; 656 defects are detected. Compared with the results of manual discrimination, the accuracy rate by using this system reaches 94.3%, and the recall rate reaches 98.7%. This system integrates the process of pipeline defect detection, reduces labor costs, and improves drainage pipeline defects'' detection efficiency, which has engineering practicability.