Research Of Recognition For Indistinguishable License Plate

The license plate recognition system is the most important part of Traffic Management System.The research on license plate recognition has been carried out for many years and has made an achievement.The use of license plate recognition systems has been widespread in parking lots and toll stations.However,most of the existing license plate recognition systems can only work well in simple scenarios.When the plate is difficult to recognize in the pictures,the accuracy will be low.This paper is devoted to studying the license plate recognition algorithm under this condition,and designing the license plate detection and license plate character sequence recognition algorithm,based on deep learning,which can accurately recognize the plate even under complex conditions.This paper analyzes the reasons for the low accuracy of the recognization of low quality pictures,which can be summarized into two points.First,the license plate is excessively tilted due to angle problems;second,the license plate is blurred due to the weather,motion blur,too far away,or defaced license plates.It can be seen as noise added to clear pictures.This article focuses on improving the accuracy of the recognition of low quality license plate from two aspects: detecting the four vertices of the license plate,using perspective transformation to correct the license plate,and reducing the interference of the tilt plate.Designing a license plate character sequence recognition algorithm based on deep learning with the image noise reduction module,which filters the image noise and improves the algorithm's ability of recognizing license plate.1.An improved license plate detection algorithm based on RetinaNet is designed.Besides predicting the rectangular license plate,the four vertex coordinates of the license plate are also predicted.The shape of the tilted license plate in the image is nonrectangular.The prediction of conventional license plate detection algorithm includes the license plate area and the background.Also,the irregular license plate will also affect the recognition.By predicting the coordinates of the four vertices of the license plate,the license plate area can be corrected,and thus benefits to the license plate recognition.In addition,by using mobleNet as a feature extraction network,the model's complexity is reduced,and the detection speed is improved.The clustering algorithm is used to analyze the data set to redesign the default box,in order to improve the accuracy of the algorithm.2.A image noise reduction network is designed based on RED-Net.Before the conventional character sequence recognition module,a noise reduction module is used to reconstruct the picture's information and filter noise factors which may influence the recognition.A character sequence recognition algorithm based on CTC Loss is designed.A light convolutional network model was used to fuse global features to recognize license plates without segmenting characters.