Application And Research Of Image Recognition System Of Locomotive Vehicle Number

With the rapid development of our country railway industry and the upsurge of internet information construction,information management system of the locomotive is imperative,the automatic positioning and recognition of locomotive number is an important part of locomotive information construction.The manual reading and checking of the locomotive number information requires a lot of manpower and material resources,while there is still a large amount of errors.With the progress of information technology,RFID has become the mainstream locomotive number identification method,but it has the disadvantages of maintenance cost and error rate.In recent years,in order to realize the information management of locomotives and the efficient identification of vehicle numbers,the automatic identification of vehicle number based on image processing is the research focus of the current railway locomotive identification.Automatic locomotive number recognition based on image processing is to obtain locomotive image through camera and carry out real-time automatic identification of vehicle number,including locomotive number positioning,character segmentation and character recognition.However,due to the locomotive body and external factors,the automatic identification of locomotive numbers are still great challenge.First,there are many types of locomotives,and the length of each locomotive model is inconsistent.Second the shading of other objects such as branches of trees and other objects can form a line or shadow,reflections shadow and the locomotive speed to form a virtual shadow.Third,the character itself may be incomplete due to paint shedding or fading,camera flash,dim characters in overcast weather,and special break characters.In view of the above three difficulties and analysis the advantages and shortcomings of traditional character segmentation method,This paper proposes using traditional positioning technology to do vehicle number positioning,based on convolutional neural network to end-to-end recognition of binary number image,remove the steps of character segmentation,thereby reducing the error of character segmentation.The main research contents and work are as follows:(1)Pre-processing of locomotive image: Image preprocessing is a very important part in the entire recognition process,Because there are a lot of noise in thelocomotive image acquired through the camera,Therefore,this paper firstly denotes the image of the vehicle by gaussian filter,and then presents the adaptive grayscale method in the process of image gray-scale,This method adopts the one-dimensional histogram identification method to distinguish the color of the body color and the non-blue base in the color of the body color and character color of different models,The S-channel in the HSV is acquired as a grayscale image for the blue-bottom vehicle number image,and the RGB weighted method is used for the grayscale image for the non-blue-bottom vehicle number image.(2)Locomotive image positioning: For the locomotive image,the characters in the night appear dim,and the first-order sobel edge detection has fewer feature points,Therefore,first-order and second-order sobel-weighted edge detection is proposed,On the basis of the first-order edge,a double edge can be added by second-order sobel detection,so that the quality of the edge detection can be improved.Secondly,for the red car with blue background,it combines the red channel to enhance the edge to improve the detection quality.Finally,the number of the candidate locomotive numbers can be found by the way of morphology and external contour detection,for multiple candidate numbers,trained the SVM number identification model based on the HOG character of the image of the locomotive number,the experiments show that the above method can accurately locate the locomotive number.(3)Locomotive number binarization and tilt correction:According to the difference between the background color and the character color of the positioned locomotive number image,the paper proposes a graying method for different background images,Then the morphology of the top hat operation to remove noise such as shadows and reflections,In order to adapt to the inconsistency of the partial brightness of the locomotive image,and a segmented adaptive binarization method was proposed,As the ShaoShan locomotive image would appear white background and black characters on the binary background,Therefore,the linear classifier of SVM,which is characterized by row and row histogram,is trained to reverse the background color and character color of the car number binary image.Finally,the horizontal tilt correction of the car number image is performed through the Hough transform,Experiments show that this algorithm can accurately achieve binarization and tilt correction for different models.(4)the end-to-end locomotive number identification:On the basis of positioningthis paper presents the identification of locomotive number end-to-end based on convolutional neural network,considering the number of locomotive number,marking data and training time,A convolutional neural network with four convolution layers and two fully connected layers is designed.The experimental results show that the end-to-end identification based on deep learning can avoid the problem of character segmentation and improve the recognition accuracy effectively.