Insulator Target Recognition Method Based On Parallel Imaging

As an important equipment on the transmission line,insulator is the focus of the inspection process.The automatic identification of insulators is the basis of the follow-up task of the electric power automatic inspection platform.In the process of insulator target recognition using convolution neural network,the problems of insufficient training sets,uneven categories and poor diversity become the main factors that limit the effectiveness of the algorithm.At present,the performance improvement methods of neural network on the data side can be summarized by the parallel image framework,but the process is independent of the characteristics of the target,which is a universal method suitable for public data sets.Based on the theory of parallel vision,this paper uses the compensation method to optimize the artificial scene image,mining the aerial insulator target features to the greatest extent,so as to improve the accuracy of the recognition algorithm.In order to solve the problem of data imbalance between training sets,this paper compares the up sampling effects of back propagation,deconvolution and guided back propagation on insulator features,and proposes an artificial scene image compensation method based on guided back propagation based on the feature extraction process of deep convolution neural network and the visual interpretation process of neural network,aiming at human The image quality of workshop scene is optimized by feedback.Firstly,according to the national standard of transmission line equipment,3D modeling software is used to model the insulator,and the artificial scene insulator image is created to form a parallel image database with the real insulator image.Then,the insulator features extracted by deep convolution neural network are guided to back propagation,and the image matrix obtained by up sampling is visualized and activated to optimize the quality of artificial scene image.Finally,the optimized insulator image database is extended to be the training set of convolutional neural network.The light convolution neural network with different depth and structure is trained iteratively to compare the influence of this method on the performance of the network.Compared with the neural network without parallel image training,the proposed method can improve the insulator recognition accuracy by 1.5%,recall by 1% and AUC by 2% on average without consuming more computing resources.The experimental results show that the proposed artificial scene image compensation method based on guided back propagation can further improve the insulator recognition performance of CNN.