Insulator Self-explosion Defect Detection Based On Fine-grained Classification In Aerial Images

With the development of drone technology,inspecting transmission lines by drones is gaining popularity and application.Faced with the massive picture data obtained by drones,manual defect detection is inefficient.Therefore,using computer vision technology to detect defects in aerial images has become a hot research issue.In this thesis,the defect detection of insulator self-explosion is taken as the research object,and the problem is transformed into the fine-grained image categorization of picture blocks of aerial pictures.Based on the analysis of insulator characteristics in aerial images,classification methods based on deep learning are proposed.The main work of the thesis is as follows:1.In order to solve the problem of the scarcity of samples of self-explosion insulator,based on the characteristics of real self-explosion insulators,two methods are used to do the data augmentation for the subcategory of self-explosion insulator.After that,the aerial images are cropped into image blocks,and after preprocessing and data augmentation,the data set is established.2.A multi-task learning model for self-explosion defect detection task is proposed.The three-category classification of "background,normal insulators and self-explosion insulators" is set to the main task.Due to the large differences between insulator pictures shot at different view-points,the task of classifying insulators according to shooting angles is used as an auxiliary task to enhance the generalization ability of the network.The experiment result shows that the recall rate of subcategory of self-explosion insulator based on the multi-task learning model is 0.9425.3.A hierarchical classification model for self-explosion defect detection tasks is proposed.Based on the label tree of the aerial picture,a hierarchical classification model is set up,and the multi-task learning algorithm is applied in the model.The hierarchical classification model uses a dedicated convolutional neural network to distinguish self-explosion insulators from normal insulators to increase classification accuracy.The experiment result shows that the recall rate of subcategory of self-explosion insulator based on hierarchical classification model is 0.9587.In summary,through in-depth analysis of the characteristics of aerial insulator images,the thesis proposes two types of insulator self-explosion defect detection models based on fine-grained image categorization that achieve high classification recall rate for the subcategory of self-explosion insulator.