| With the acceleration of urbanization,the scale of urban underground drainage pipelines in China is expanding rapidly.However,the level of maintenance and management of the matching drainage pipelines is lagging behind,which may cause many security risks.The use of efficient pipeline detection methods is the basis and prerequisite to improving the maintenance and management capacity of drainage pipelines.However,the most widely used closed-circuit TV detection(CCTV)in drainage pipeline detection is affected by the subjective factors of the inspectors,and the degree of intelligence needs to be improved.Combining convolution neural network deep learning algorithms to optimize the CCTV method is of great significance in improving the level of drainage pipeline maintenance and management.In this paper,the defect detection of CCTV drainage pipeline is divided into two tasks: image defect identification and defect classification.The following work is done:First,319 inspection reports of CCTV census projects for urban rainwater and sewage drainage pipelines were collected.Pictures of drainage pipelines in the report were sorted out,and experimental datasets were set up,which included 11 kinds of defective pipeline images and normal pipeline images.A total of 7152 image samples were collected,and image samples were pre-processed for black edge removal,size adjustment,image noise reduction,and brightness balance.Then,for the CCTV drainage pipeline defect identification task,considering the drainage pipeline image characteristics from a large range to a small range and diverse distribution,the Shuffle Net is structurally improved.By adding a convolution calculation branch and expanding the network width,the Shuffle Net B network model is designed.The image defect-recognition experiment of the CCTV drainage pipeline using the Shuffle Net B network is carried out.The classification performance of the Shuffle Net B network is verified by experiments.The results show that the improved Shuffle Net B model has better classification performance,with a classification accuracy of 98.2% and 2.1%higher than Shuffle Net.Finally,in view of the CCTV drainage pipeline defect classification task,considering the limited number of samples and the unbalanced distribution of categories,three specific optimizations are made based on the SHUffle Net B network model: A migration learning method is proposed to migrate some image feature information from the Shuffle Net B drainage pipeline defectrecognition task to the Shuffle Net B pipeline defect classification task;Use an image amplification method that guarantees sample quality.The weight-adjusted crossentropy loss function is used to adjust the loss to calculate the bias according to the number of samples in each defect category.The experimental results show that the Shuffle Net B network model optimized by multi-classification has better classification performance in the pipeline defect classification experiment,and it can capture the pipeline defect characteristics very well.The classification accuracy of 11 types of drainage pipeline defects reaches 88%,and the single-category recall rate reaches 93.3%. |
