Research On License Plate Recognition System Based On YOLO

With the growth of computer computing power,artificial intelligence technology can be well applied,the traditional field gradually toward the direction of intelligent development.In2016,General Secretary Xi Jinping emphasized the construction of a new smart city in the collective learning.Among them,the license plate recognition field in smart traffic is developing in full swing,and the license plate location algorithm based on deep learning constantly impacts the traditional algorithm.The purpose of this paper is to improve the existing YOLOv3 algorithm and verify it as an application-level algorithm that highly fits the application scenarios of license plate location recognition.The algorithm improves the recognition speed on the premise of ensuring the recognition accuracy,so as to meet the requirements of high real-time and high accuracy application scenarios.The main work of this paper is as follows:(1)Design two schemes to apply YOLOv3 algorithm to license plate location recognition system.In the first scheme,YOLOv3 algorithm is only responsible for selecting the position of license plate,and then intercepting the license plate area using the traditional method for character recognition.In scheme 2,YOLOv3 algorithm selects license plate and license plate characters at the same time,and uses the spatial distribution characteristics of license plate elements to recognize characters.Relevant experiments were designed,high real-time performance was taken as the standard,and scheme 2 was finally selected as the basic research direction of this paper.(2)Based on scheme 2,the following optimization of YOLOv3 algorithm is made for the scene of license plate location recognition.Before the model training,the k-means ++ algorithm is used to select the initial anchor points of license plates to solve the slow model training phenomenon caused by improper selection of initial anchor points.In the process of multi-scale information fusion,this paper puts forward a detail enhanced auxiliary network: in feature extraction adding level of all the connection layer after layer,and keep the direction of propagation constant,the model to extract image detail is greater than the original model,license plate character recognition effect,improve small target to vision character is too small in license plate recognition problem has larger ascension.In the regression process of Loss,the size misalignment penalty term is introduced,so that the model has a sharp decrease in the confidence of the prediction box that does not conform to the size of license plate features,and part of the calculation pressure is shared in the regression process of the model.Although this operation will increase the volume of weight files,it can improve the recognition speed and accuracy at the same time.(3)In order to evaluate the performance of the model more comprehensively,the open source data set VOC2019 was used for training and verification,with mean Accuracy Percentage(m AP),recognition Accuracy and frame rate as evaluation criteria.The optimized YOLOv3 algorithm was compared with Easy PR(traditional algorithm),and the m AP of optimized YOLOv3 reached 97.5%.Compared with Easy PR algorithm,the frame rate of optimized YOLOv3 reached 44.7628 FPS,3.75 times higher than that of Easy PR algorithm,but the recognition accuracy decreased by 1%.And in different application scenarios(such as daytime,occlusion,etc.),the optimized YOLOv3 algorithm has better recognition stability.Compared with the original YOLOv3 algorithm,the recognition accuracy is improved by 3.7%.The improved YOLOv3 algorithm can accurately recognize long-distance vehicles,license plate wear,occlusion,blur,complex Chinese characters,and has feasibility and application value for high real-time,high accuracy,and high fault tolerance scenarios.