Research On Insulator Defect Detection Based On Faster R-CNN

With the continuous development of China’s economic level,the power industry plays an increasingly important role,and with the continuous extension of transmission lines,the power grid has higher requirements for the safe operation of insulators.Therefore,better insulator defect detection is an important part of the stable power system.At present,the insulator inspection method has basically realized the UAV inspection instead of manual inspection,but due to the special characteristics of insulators themselves,the background of the images taken by UAV is complicated and the defect targets are small,which increases the difficulty of target detection,and there are certain disadvantages of using traditional detection methods,therefore,this thesis combines deep learning technology,and analyzes and compares a variety of algorithms through in-depth research on deep learning framework.After that,an insulator defect detection algorithm based on Faster R-CNN is proposed,and the main research contents are as follows:(1)Since the batch normalization method is not effective when the batch size of Faster R-CNN can only be 1,replace the normalization method from batch normalization to group normalization,which is almost independent of batch size,to improve the algorithm normalization effect and reduce the error rate.(2)Because of the limited performance of the feature extraction network used in the Faster R-CNN,to improve the feature extraction capability of the algorithm,the feature network uses a deeper Res Net-50 network structure with residual neural networks to enhance the feature extraction capability of the algorithm and improve the detection accuracy.(3)To further improve the detection performance of the algorithm,the channel attention mechanism ECA is introduced to achieve cross-channel interaction and improve the feature extraction network performance without increasing the complexity.(4)The performance of the algorithm is again improved by improving the non-maximum suppression algorithm using Soft-NMS to solve the problem of missed detection caused by the close alignment of targets in the non-maximum suppression method.In addition,the initial images are processed by various data enhancement methods to obtain the experimental dataset and manually labeled.By comparing the experimental results and the effect of complex sample detection before and after the algorithm improvement,we can conclude that the detection performance of the improved algorithm is significantly improved,and the average accuracy is improved by 14.7% and the F1 value is improved by 14.3% compared with the traditional Faster R-CNN.In summary,the improved Faster R-CNN can better achieve insulator defect detection and provide a new idea for transmission line insulator fault inspection.