Research On Crack Classification Method Of Underground Drainage Pipe Based On Deep Learning

As the common means of gas and liquid transmission,pipelines have been widely used in underground sewer systems.Pipelines need to be inspected and maintained regularly.Otherwise,serious pipeline defects may endanger public safety and require costly system repairs.In engineering applications,although pipe scan data is obtained by automated technologies like CCTV or SSET,defect diagnoses are still performed by hand.Artificial judgments are subjective and can lead to errors.Therefore,this thesis carries out the research on the crack classification method of underground drainage pipelines based on deep learning.The research content of this thesis is as follows:(1)Construction of underground pipeline crack dataset.According to the sewer condition assessment criteria,PACP(Pipeline Assessment Certification Program)requires encoding pipe cracks at a sub-category level.Therefore,in cooperation with the Oklahoma State University,the CCTV video images of underground pipeline cracks were collected and sorted out in this thesis,and some pre-processes were carried out on them,such as image labeling and blurring information label boxes.In addition,to solve the problem of unbalanced sample data of pipeline defect types,data enhancement is adopted to expand the crack sample set and achieve data balance.Finally,the pipeline crack data set is constructed as the training data of this thesis.(2)Mask-guided attention for classification algorithm is proposed.In this algorithm,a collaborative learning model based on image segmentation and classification is studied,and a model based on convolutional neural network is constructed by combining pixel-level segmentation and global image-level classification annotation.The method consists of four modules: feature extraction module,multi-scale feature fusion module,mask-guided attention module,and sub-category classification module.Firstly,VGG-16 is used as the backbone to fully extract features by combining the multi-scale feature fusion method.Secondly,mask is used as an attention mechanism to refine the features,which enhances the feature representation of cracked pixels,thus reducing the ambiguity of the crack fine-grained classification task.Finally,the sub-category classification results are obtained based on the feature map of background clutter elimination.(3)A lightweight pipeline crack classification algorithm is proposed.In laboratory scenarios where model volume is not considered,the algorithm proposed in(2)can obtain higher accuracy.However,the above algorithms increase the complexity of the model inevitably while further improving the accuracy.Therefore,this thesis considers introducing knowledge distillation algorithm to compress the model.Taking the fine-tuned Res Net50 as the teacher model,a lightweight network CMI(Combined Mobile Net V2&Inception)for crack classification was designed as the student model for distillation,aiming at meeting the needs of the deployment of the edge end.The experimental results show that the classification accuracy is increased by 1.41%,and the F1-score is increased by 1.12%.To sum up,the above two methods carry out research respectively from the perspective of classification accuracy and model lightweight to obtain a model that can meet the needs of practical applications.The proposed method achieves good classification performance on the pipeline crack dataset,and its experimental results are better than the algorithms based on deep learning,which lays a solid foundation for the development of automatic PACP sewage pipe detection system,and further promotes the digital and intelligent process in this field.