Key Technology Research On Contact Network Mast Number Plate Identification

With the rapid development of railroads in China,the transport volume and operating hours of trains are increasing year by year,while the window time for maintenance operations is limited.As a key component of the power supply system of the electrified railroad,contact network maintenance is the main content of the window operation of the power supply section,and it is especially important to carry out maintenance operations efficiently.The maintenance and repair of the contact network system is not only to know what was wrong,but also to know where is the problem located.There is a contact network mast at each 50 meters along the railroad line,and the mast number,which uniquely represents the mast location information,is attached to the mast.By identifying the numerical number of the plate,the location of the catenary wire can be found,it is easy for staff to quickly locate the operating area.This study develops and optimizes the mast number plate algorithm for the automatic identification of railroad contact network mast number plates in the traditional image processing method,and also explores the feasible method of plate identification using deep learning theory.This study first investigates the method of railroad mast number plate recognition by the traditional image processing algorithm based on image features.The traditional image processing algorithm contains three parts to recognize the contact network mast number plate number plate positioning,character segmentation and character recognition.This paper proposes a number plate localization method using character feature information Euclidean distance threshold,firstly extracts the edge of the image using edge detection,then finds the straight line of the edge of the contact network mast using Hough straight line detection and finds its deflection angle relative to the vertical direction,and then corrects the image by rotating and transforming the image with this angle.The corrected image is then binarized using a modified adaptive multi-threshold binarization algorithm,and about 90% of the interfering connected domains are filtered out by multi-featured connected domain screening.Finally,the remaining small number of contiguous domains are precisely found using a new method of Euclidean distance thresholding algorithm,so that the location of the number plate can be precisely located.The second part of the character segmentation method uses a common projection segmentation,and finally a template matching algorithm is used to identify each character.The algorithm was tested in the existing data set to achieve 95%recognition accuracy.Traditional image processing methods is fast and effective to detect and identify contact network mast number plates but not robust enough for various scenarios.Deep learning based target detection algorithms for the same task was explored.In this paper,the YOLOv2 detection algorithm based on the Res Net50 base network is chosen to achieve end-to-end recognition of railroad contact mast plates.The first step is the localization of the number plate,and the first single-class YOLOv2 network is trained to position the mast number plate.The second step is the recognition of the mast number plate,training a YOLOv2 recognition model that divides 10 Arabic numbers into 10 categories.Finally,the mast number plate localization model and the mast number plate recognition model are linked together as a key part of the contact network mast number plate recognition,and a 96% recognition accuracy is achieved after testing the existing data set using this method.