Research On Detecting Method Of Sewer Defects Based On Vision

With the deepening of urbanization,the scale of urban underground pipelines is expanding.Regular inspection and maintenance of underground pipelines can greatly reduce the social and economic damage caused by pipeline defects.At present,CCTV(Closed Circuit Television)detection is the mainstream underground pipeline detection method.The underground pipeline robot is used for live shooting of the pipeline,and then the videos are judged by professional inspectors.A large number of pipeline videos are detected manually,which has the disadvantages of fatigue,subjectivity and inefficiency.In order to solve the above shortcomings,an integrated intelligent detection method for pipeline video based on computer vision is proposed.This method can automatically determine whether the pipeline has defects,what defects exist and where the defects occur.Pipeline defect detection is divided into four steps: video frame anomaly detection,abnormal frame defect classification,pseudo anomaly frame filtering and defect information recognition.Compared the performance of support vector machine model based on HOG(Histogram of oriented gradients)feature,extremely random tree model based on GIST(Gabor filter features of multi-direction and multi-scale)feature and convolutional neural network,and finally chose convolutional neural network as the detection model.Aiming at the problem of poor real-time performance of large-scale deep network algorithms,depthwise separable convolution and group shuffle convolution are used to replace traditional convolution,which greatly improves the detection efficiency of the model while maintaining accuracy.Aiming at the problem of high over-detection rate due to pipeline topography and artificial yaw control,the improved bidirectional optical flow method is used to filter pseudo anomaly frame.Aiming at the problem of low accuracy in actual detection,the transfer learning and model snapshot integration strategy is adopted to improve the accuracy.At the same time,video character detection and recognition algorithm is designed to meet the functional requirements of automatically recording information such as defect location and distance while finding the defect.According to the video character characteristics of pipeline scenes,the maximum stable extremum region algorithm and the non-character region filtering algorithm based on cascaded classifier are used for character localization.The hierarchical clustering algorithm based on color and scale information is used for character segmentation.Template matching and OCR software are used for number and text recognition respectively.Compared with the traditional detection method,the proposed integrated intelligent detection method for pipeline video has greatly improved the accuracy and actual scene adaptability.At the same time,the use of anomaly detection for video defect segment extraction,combined with the manual re-inspection can shorten the detection time of professionals by about 70% while ensuring a higher defect detection rate.