Design Of Number Plate Real-time Recognition System

New technologies and industries have been rising up in public life in recent years,among which artificial intelligence should be paid more attentions for rapid progress and changing people's customs.Automatic number plate recognition(ANPR)is an essential part of the intelligent traffic system which is a real-time,accurate and efficient traffic and traportation management system.At present,number plate is the only identification of the car,for which ANPR is fundamental to realize vehicles' intelligent management.Research for ANPR has the practical significance for its necessity in vehicle control,traffic monitor,vehicle trace analysis and other application scenarios.At present,there are many traditional approaches that have achieved a rather high accuracy to solve ANPR,almost all of which are separated into three steps of localization,segmentation and recognition.However,these approaches,especially in segmentation progress,are limited to some specific conditions including light intensity,orientations,rotation and distortion angle of plates,etc.Method in this paper is based on convolutional neural network and it can integrate segmentation and recognition process to decrease the influence of environment.In general,method to solve ANPR can be divided into two stage:localization and recognition.For localization,cascade network detection now is a relatively common and advanced alogrithm,whose technique is to filtrate and shrink target area by neural network and then choose the location of the target by NMS algorithm.In some extend,cascade network decreases the sensibility to illumination and number plate angle,which is a problem for traditional method based on graphic analysis.But it's limited in high resolution ratio graph beacause of the using of slide window.In this paper,we will introduce the detection method based on Yolo model,which is more excellent than cascade network in dealing with complex environment and high resolution ratio graph.For recognition,character segmentation and character recognition are two stage to solve the problem in traditional method.Beacause of the limitation of character amount and character form,traditional recognition method can achieve a high accuracy under accurate segmentation method.However,it's hard to implement an accurate segmentation under some unfavorable circumstances such as underexposure and number plate distirtion.In some cases such as parking lot,the method including segmentation can perform well beacause it's easy to capture favorable graph,which is difficult in more complicated cases such as drone camera and traffic monitor.In this paper,distinct from traditional approaches,a network including a convolutional neural network(CNN)that operates directly on the image pixels is employed as a substitute of the integration of segmentation and recognition.Models in this paper are designed and trained under Darknet and Tensorflow framework Also,based on CNN model we proposed,we developed a C/S-architecture software to supply recognition function for videos,graphs and real-time video stream.The software will only execute the inference process of the model and feedback recognition results.