Research On 4C Defect Recognition Method Of Railway Contact Network

The rapid development of China's high-speed railways has increased the pressure of railway bureaus on the inspection and maintenance of infrastructure along the line.As an important part of infrastructure,the contact network is used frequently and without backup.Its integrity and stability directly affect the safety of trains and personnel.At present,the dynamic detection of the contact network uses an on-board camera 4C device to capture images and uses traditional low-precision computer vision algorithms.However,this type of algorithm has low detection effect and is not conducive to large-scale data set detection.In recent years,the rise of deep learning has made the accuracy of intelligent recognition quickly surpass the naked eye and traditional detection algorithms.Therefore,this paper introduces the method based on deep learning into the defect detection of the contact network,and does the following work:Firstly,it analyzes the advantages and disadvantages of the current contact network recognition technology,as well as the development status of deep detection-based intelligent detection algorithms.Aiming at the original data set of the contact network,a series of problems such as small defects,similar components,and uneven distribution of directions,the technical process of this experiment was designed.The technical process is mainly divided into two stages to achieve,the first stage is to locate the contact network components,through the analysis of Opencv correlation matching method for image matching,based on deep learning characteristics and data characteristics,using deep learning based R2 CNN algorithm Target object recognition with tilt angle.After training and testing operations,the final result shows that the algorithm has a good detection effect,and can continue to the next stage of research.The main purpose of the second stage is to intelligently identify component defects.The source of its data set is cut from the detection frame of the first stage component identification.By analyzing the development and advantages and disadvantages of the target detection algorithm,the Faster R-CNN algorithm based on Res Net extraction network was initially selected for defect identification.Through training and testing the algorithm,it has achieved certain results,but it can not satisfy the needs of high recall well,and there is a problem of slow recognition speed.In order to obtain a better detection effect,the RFCN algorithm based on Res Net network is usedto identify defects,eliminating the time-consuming operation of full connection for each area of interest.Compared with Faster R-CNN detection algorithm,the detection result of this algorithm has better improvement than Faster R-CNN detection algorithm,but its detection effect still can not meet the needs of efficient detection of large-scale contact network data sets.Therefore,the multi-scale convolution detection algorithm Refine Det was used again for the experiment,which combines the advantages of SSD algorithm and Faster R-CNN algorithm.It can be seen from the comparative study that this algorithm improves the detection efficiency of this contact network dataset better than existing algorithms.Finally,corresponding improvement schemes are designed for existing algorithms and data problems.By analyzing the image problems of the data set,preprocessing operations such as sharpening and homomorphic filtering are used.According to the experimental results,the filtering has a certain improvement effect on some components of this data set;in addition,according to the defect problems after component testing Bounding box regression loss and soft NMS improvements.Through experimental comparison,it is judged whether the algorithm is feasible for this data set.The comprehensive analysis and comparison shows that the research method adopted in this paper is effective for the intelligent defect identification of the contact network data set,but there are still some problems,and the algorithm needs to be continuously strengthened to quickly and accurately improve the efficiency of railway contact network defect identification.