Research On Image Classification Of Small Sample Power Inspection Based On Self-supervision

Power inspection is to inspect and detect the operation state of transmission line and its auxiliary equipment.Its main purpose is to check the defects and hidden dangers of equipment and ensure the normal operation of transmission lines.With the smart grid policy proposed,it is particularly urgent to make the power inspection smart.However,the power inspection mainly based on manpower and partially based on smart devices at present.The problem of relying on manpower to audit defect and hidden danger images has not been solved.In recent years,deep learning has been widely used in the field of computer vision.Deep learning can analyze and obtain the internal law of image or video,so the effect of image recognition is far more than that of traditional visual technology.Therefore,this paper focuses on the problem of defect and hidden danger image classification in the scene of power inspection by using deep learning.This paper takes a few-shot image classification as the research object,analyses the characteristics and classification difficulties of defect and hidden danger image deeply,and puts forward a few-shot defect and hidden danger image classification network based on self-supervised to achieve intelligent auditing image.The work of this paper is as follows:(1)This paper proposed a few-shot defect and hidden danger image classification model based on Arcface.At present,there is no public dataset of defect and hidden danger for research.This paper filters the sample of defect and hidden danger image and builds the defect and hidden danger dataset.This paper solves the problem of fewer samples by using few-shot learning.Based on the characteristics of potential defects,this paper improves the loss function of prototype networks and increases the accuracy of classification effectively.The experimental results show that the accuracy of using 5-way1-shot and 5-way5-shot in the defect and hidden danger dataset is improved by 8.13%and 8.52% respectively.(2)In order to improve the generalization ability of the network,this paper adds the pre-training stage and proposes a self-supervised model based on masked autoencoder.The model uses a self-attention structure from transformer.After pre-training on the large dataset Imagenet,the model is transferred to the downstream task to fine tune,that is,the few-shot classification task.The experimental results show that the accuracy of the selfsupervised model based on masked autoencoder proposed in this paper can reach 83.92%,which is better than most of the existing self-supervised models.(3)Combined with the few-shot defect and hidden danger image classification model based on Arcface and the self-supervised model based on masked autoencoder,a few-shot defect and hidden danger image classification network based on self-supervised is proposed in this paper.The accuracy rate on 5-way 1-shot is 78.44%,which is 13.13%higher than that of the prototype network;The accuracy rate on 5-way 5-shot is 92.60%,which is 11.87% higher than that of the prototype network.This method can obtain more robust high-level semantic representation,and then improve the accuracy of classification.Aiming at the challenging task of defect and hidden danger image classification,it solves the problems of fewer samples and relying on manual labeling,and improves the efficiency and accuracy of power inspection.