Transmission Line Bolt Defect Classification Method Based On Feature Transfer Image Super Resolution

Due to the long-term operation of the transmission line in the field,the bolts on the transmission line are affected by the environment and their own tension,blots are prone to defects such as missing pins,loose nuts,and missing nuts,which affect the safe operation of the transmission line.Therefore,it is necessary to carry out intelligent detection of transmission line,effectively identify bolt defects and eliminate hidden dangers in real time.However,due to the small size of the bolt,the images collected by aerial photography are prone to blur,and the direct defect classification is not effective.Therefore,this paper proposes a secondary processing method,which first performs image super-resolution processing on the bolt and then performs defect classification.Aiming at the problems of blurring and low resolution of bolt images collected during inspection of transmission lines,according to the characteristics of high similarity between bolts,a super resolution processing method of bolt images based on adaptive feature transfer is proposed.This method is the first time to introduce feature transfer into bolt image super resolution,first compare the feature regions of low resolution images and reference images,and transfer the features of the regions with high similarity between the images,and propose a similarity correction module.Adjust the scale of the transferred features according to the similarity of the features.Experiments compare the results of bolt image super resolution with different super resolution models,and the results show that the bolt super resolution image of this method is clearer,and the PSNR and SSIM indicators are better than others.In order to further improve the image quality,this paper adds the adversarial loss function and the perceptual loss function to the original loss function,and adds the similarity constraint of the transfer feature to the perceptual loss function to ensure the accuracy of the transfer feature.The experiment compares the results of bolt images under different loss functions.The experimental results show that the image visual effect is better under the loss function constructed in this paper;The defect recognition accuracy of 3.61% was improved.The effectiveness of this method is fully verified.In order to better extract bolt surface features,this paper proposes a joint double attention bolt defect classification method with WRN as the backbone network.Firstly,the spatial dimension of the feature map is compressed by SENet attention to obtain the global features of the channel dimension,enhances the network’s attention to important information in a weighted manner;and then embeds collaborative attention,decomposes the enhanced feature map into two one-dimensional feature vectors,and establishes long-term dependencies in one spatial direction,saves accurate information in the other direction.Location information helps the network to more accurately extract key feature information,thereby improving the accuracy of classification.The experimental results show that the bolt classification accuracy of this method is improved by 1.26% compared with before embedding attention.