Research On Multi-label Classification Method For Sewer Defects Based On Graph Convolution

With the development of modern cities,the sewer pipes buried deep underground in the early days have been overwhelmed,and the types of sewer defects have become complex and diverse.The traditional detection methods relying on manpower can no longer meet the detection needs of sewer pipes.Moreover,due to the differences in image acquisition equipment and uncontrollable factors in the internal environment of the sewer pipes,the image quality of a large number of sewer pipes is low and the details are seriously lost,making it difficult to identfy defect features and affecting the detection effect of sewer defects.Therefore,it has important academic research and application value to carry out research on intelligent identification of sewer defects.The existing sewer defect classification methods mainly include traditional image classification methods and image classification methods based on deep learning,but most o f these methods cannot effectively classify sewer images with a large number of defect lables;The difficulty of mining correlation information and the difficulty of extracting fine-grained features of images exacerbate the difficulty of mulfi-label classification of sewer defects.According to the above analysis,this thesis focuses on the multi-label classification of sewer defects.The main research contents and results are as follows:(1)In view of the fact that the existing classification methods do not have the ability to deal with multi-label data and the image data imbalance of sewer pipes,this thesis proposes a twostage sewer pipe defect classification method based on graph convolutional network.In the first stage,a shallow convolutional neural network is designed to classify normal pipelines and defective pipelines;in the second stage,a multi-label classification model based on graph convolutional neural network is designed,which can classify the correlation between labels The information is modeled,and the label relationship is updated through the graph convolutional network,thereby enhancing the multi-label classification ability of the model.In order to test the classification performance of the first stage,evaluation experiments were carried out on the sewer dataset,and the results show that the classification effect of the proposed shallow convolutional neural network is improved compared with other shallow convolutional neural networks.(2)In view of the fact that the existing multi-label classification method based on graph convolution ignores the potential high-order correlation problem in multi-label images,this thesis proposes an adaptive graph convolution network that integrates global and local regional features,which can be obtained from different feature extraction.Firstly,the local features of image defects are extracted using the region prediction algorithm;then,the dynamic adjacency matrix is constructed in combination with the global feature extraction,and the label correlation information is adaptively updated through the graph convolution algorithm,so as to complete the multi-label classification task of defects.The analysis of experimental comparison results with other advanced methods shows that the algorithm proposed in this thesis can achieve more accurate multi-label classification results on the sewer dataset,and the recognition effect on individual difficult labels has been significantly improved.(3)For the problem that existing multi-label image classification models lack strong finegrained feature extraction capabilities,this thesis proposes a multi-label classification method based on triple attention Transformer and static graph convolution.First,integrate the information of the label itself by introducing category attention;then use category attention features to enhance the ability of label embedding representation,further combine self-attention and crossattention to build a Transformer model,and model the correlation information between labels to represent Long-distance dependencies between lables;Finally,the representation of lable neighborhood information is enhanced by means of static graph convolution,which improves the model’s ability to extract fine-grained features and realizes multi-label classification of sewer defects.The comparative experimental results show that the proposed method has a good classification effect on defect images containing fine-grained features,and performs excellent in various evaluation indicators.