Semantic Segmentation Of Catenary Monitoring Image With Retrieval And Location Of Key Devices

Catenary is the key device to obtaining electric energy for powered car train-set.The development of using intelligent detecting methods to detect the catenary’s running state which is an effective measuret to maintain the high-speed railway’s normal working situation.And this method has a promising applicant future and profound meanings.This paper studies image semantic segmentation algorithm based on deep learning.The aim is to macroscopic analysis and understanding the catenary monitoring scene content,retrieval and localization on the key devices of the catenary system,improve the efficiency of the operation state detection of catenary.This paper takes catenary monitoring images of CCLM as the research object,proposes a feature attention based bilateral segmentation network for catenary monitoring image,I call it as CMI-Bi Se Net.The CMI-Bi Se Net achieves the semantic segmentation of catenary monitoring images which have complex scenes and get from different environment condition.At the same time,because the current image semantic segmentation network failed to balance the accuracy and speed aspect,it studies to improve the speed of semantic segmentation on the foundation that segmentation result has a good accuracy.Take the dropper as an example,this paper also studies retrieval and localization on the key devices of the catenary system after semantic segmentation.The main work and innovations of this paper are as follows:1.Develop a tag tagging program that called pixel＿tag,based on the pixel＿tag,this paper constructs the catenary monitoring image semantic segmentation dataset.By study the data requirements of deep learning,this paper collectes the appropriate monitoring images from massive catenary monitoring data.Because the semantic segmentation is pixel-level dense image classification,this paper uses pixel＿tag manually complete the production of 300 pixellevel accuracy tag images.Then choosing the appropriate method to simultaneously augment data on original image and its semantic tag image.Finally,it gets enough data that suits the requirements of the deep learning.This work provides sufficient data for the experiment.2.This paper proposes a feature attention based bilateral segmentation network for catenary monitoring image.Aiming at the problems of semantic segmentation network model acceleration methods.This paper improvs the network based on the Bi Se Net.The Bi Se Net has spatial path and context path,and the two path use parallel computing,this method greatly improves the segmentation efficiency.And in order to ensuring the segmentation accuracy,this paper uses Res Net to reliace the lightweight network as the backbone structure of the context path.From the value of PA and MIo U,and compare the semantic segmentation map,this paper analysis the segmentation results of five different situations.At the same time,compare the segmentation speed of different networks.It prove the CMI-Bi Se Net can get good segmentation results for different catenary monitoring images.And it can effectively improve the speed of segmentation on the foundation that segmentation result has a good accuracy.3.In order to identify the situation of the in-site monitoring images of the catenary system,this paper taking the most fragile component dropper as the main research target.Based on the semantic segmentation results and use the image processing technology to complete the retrieval and positioning of the dropper.At last,get the image that only contain dropper line.This work provide the preparation for the anomaly detection of the dropper in the future.