Research On Image Defect Detection Methods For High Voltage Transmission Lines

Transmission lines system is one of the essential components of the national economy.Also,the system stability is guarantee of the population daily lives.However,the gradual expansion of the power system scale,the variety of components in the line system and the complexity of defects,which result in such as high workload and low efficiency of manual discrimination of aerial images of high-voltage transmission lines,and have high requirements for the physical and technical level of maintenance personnel.Therefore,this paper proposes a method which used for locating key components of high-voltage transmission line images and defect segmentation detection based on computer vision technology.This can promote intelligent development of transmission line inspection.The main work is as follows:(1)The method for locating key components of high voltage transmission lines based on improved YOLO v3 is proposed for aerial inspection global in this article.Firstly,the 12,000-sample transmission line inspection data set was constructed for model training and testing.Secondly,based on the morphological characteristics of the inspection object,the Darknet feature extraction network was improved by the Res2 Net feature extraction module to enhance.The fine-grained detection capability of the network,which can solve the problem of low recognition accuracy caused by the various scales and angles of key components in aerial images.The H-Swish activation function is also used to replace the original Re LU function to reduce the computational complexity of the model and ensure the inference speed of the model.Finally,the Softer NMS algorithm was introduced to replace the NMS algorithm in the YOLO v3 detector to circumvent the violent deletion of detection frames.The results show that the proposed method can achieve simultaneous and accurate positioning of five types of components(insulators,hammers,connectors,towers and signs),and two types of defects(bird’s nests and foreign objects).This system can provide an increase of 6.9% in intersection over union value and mean average precision of 91.6% compared to that before the improvement.(2)For the blurring of the cropped subgraph in the key components after localization,the adversarial network with improved Wasserstein loss function and generator structure is introduced,for the super-resolution reconstruction of blurred images.In order to discriminate the defects of the key component images after super-resolution reconstruction,the Res Net feature reuse structure is proposed based on Mask RCNN segmentation algorithm,by which the morphological features of the shallow network are reused in the deep network to fuse with the semantic features;Also,the Point-Set strong a priori anchor frame optimization mechanism is introduced to improve the defect segmentation fit.Experimental results show that the method provides a better segmentation effect on the pin loss and insulator self-burst,and the mean average precision of the optimal defect detection scheme reaches85.3%.(3)Finally,based on the trained model,a Java Web-based software system for high voltage transmission line defect detection was designed.Locally saved tower inspection images of defects can be automatically analyzed,at the same time the test report is output.This also provides convenient services for later maintenance.