Research On Deep Learning Based High-Speed Railway Catenary Image Analysis Technology

Catenary is a crucial device of electrified railway industry and its state is related to the safety and reliability of railway.Nowadays,Catenary Suspension State Detection and Monitoring Device(4C)is widely used in catenary state monitoring.4C uses high-definition industrial camera installed on detecting vehicle to collect images,and judges the state of catenary by analyzing these images.Due to the high cost,low efficiency and unstable results,mannual check has been eliminated.Intelligent image analysis technology has become the inevitable demand.In recent years,deep learning,especially deep convolutional neural network,has brought great technological innovation to computer vision,and created great application value in Internet,security and other fields.However,in catenary image analysis,the existing technology still stays in the traditional image processing algorithm and directly applied general deep learning algorithm,and lacks the in-depth research to adapt catenary.Therefore,it is an important topic to study the catenary image analysis technology based on deep learning.In this paper,insulator is selected and defect analysis of insulator based on deep learning is studied.Before analyzing the insulator defects,the insulator image should be obtained from the catenary image so an insulator detection algorithm without background result is needed.This paper proposes InsulatorDet,an accurate insulator detection algorithm.The algorithm constructs the network with MobileNetV3 and feature pyramid network.During the training,the positive and negative samples are defined by local pixels.The regression target of positive sample is defined according to the shape of insulators.Prior area enhancement module based on attention mechanism is designed as the parallel branch to enhance the important area of image.During the test,Non-Maximally Suppressed is applied on the output of each pixel to get the rectangle detection box with angle.Experimental results show that InsulatorDet outperforms other general detection algorithms in accuracy and practicability.In the analysis of insulator defects,this paper studies from two aspects: data and algorithm.At the data level,in order to solve the problem of insufficient number of training images of defective insulator,this paper proposes to use Generated Adversarial Network to expand the dataset.At the algorithm level,a defect recognition algorithm called DefectNet is proposed.Based on VGG,this algorithm proposes three improvements.First,a refined expression network is designed to constrain and recombine the column features.Secondly,the center triplet loss is proposed as an intermediate loss to optimize the feature metric space.Thirdly,self supervised learning is applied to automatically learn the local feature that can best express the image.Experimental results show that the three improvements can effectively improve the defect recognition results.