Research And Implementation Of License Plate Recognition Technology Based On Deep Learning

With the development of modern society,the number of motor vehicles is increasing,which puts forward higher requirements for intelligent traffic management.Vehicle license plate information is the main feature of vehicle,so license plate recognition algorithm plays an important role in the intelligent traffic management system.License plate recognition algorithm is widely used in entrance and exit control,toll station and other scenes.In addition,it has related applications in image recognition such as UAV and mobile phone shooting.The traditional method of license plate recognition needs a variety of methods to synthesize,and the implementation steps are complex,and there is a certain gap in recognition rate and robustness compared with the deep learning method.There are various forms of license plates in China,with complex colors and different sizes.Now there are eight character license plates.The traditional methods are difficult to recognize the license plate accurately under the conditions of illumination change,image blur,license plate damage,license plate size change and so on.This paper mainly aims at the lack of robustness of traditional methods in license plate recognition,and proposes a license plate recognition method based on deep learning,and applies it to the campus entrance and exit scene to achieve a more robust and accurate license plate recognition.The main research work and contributions of this paper are as follows:(1)This paper studies the target detection algorithm in deep learning,and applies it to license plate location through improvement.This paper mainly implements the improved Multi Task Convolutional Neural Networks(MTCNN)algorithm,the improved You Only Look Once(YOLO)algorithm,and proposes the improved LP-RetinaNet algorithm based on RetinaNet.The improved mtcnn algorithm solves the problem of image multi-scale,and can accurately and robustly realize the task of license plate location.The accuracy of this method in self built data set is 99.7%,and in Chinese City Parking Dataset(CCPD)2019,the location accuracy on the weather data set is 99.49%,which can realize the more robust license plate location.By using the high-resolution image data set training model,based on the improved YOLO algorithm,the accurate location is realized under the condition of low image resolution.This method has four main improvements: first,improve the feature extraction network;second,change the output of the feature map to 14 × 14.Thirdly,the Generalized Intersection over Union(GIoU)loss is added to the original loss function;fourthly,the calculation method of IoU in the training stage is GIoU.Through the experiment,it is found that the accuracy of the improved YOLO algorithm based on DetNet59 backbone network in self built data set is 99.97%,and the location time is only 46.7ms.The accuracy of the backbone network based on DarkNet19 in self built data set and CCPD 2019 weather data set is 99.95% and 98.44% respectively.Finally,in order to improve the precision of location and realize the correction of inclined license plate,this paper proposes a license plate location algorithm based on LP-RetinaNet,which can realize license plate location and correction through the network.Compared with the original RetinaNet algorithm,this algorithm improves the backbone network model and adds Single Stage Headless(SSH)module,adding the loss of key points,the model realizes the lightweight design,and the accuracy rate reaches 97% on the self built data set,and 99.63% on the CCPD 2019 weather data set,which realizes the more robust license plate positioning.(2)This paper studies the traditional method of license plate character segmentation algorithm,combined with the laboratory and previous research to achieve a more accurate license plate character segmentation.In this paper,the algorithm of license plate character segmentation achieved 99.6% accuracy on 500 fixed license plates,which can basically achieve a more accurate segmentation of conventional license plates.(3)The character recognition method based on the Convolutional Neural Networks(CNN)and the sequence recognition method based on the Convolutional Recurrent Neural Networks(CRNN)are studied.The more accurate license plate character recognition is realized through the improvement.In this paper,a license plate character recognition method based on the improved lenet-5 is proposed,which has an average accuracy of 99% in Chinese characters,letters and numbers.In addition,based on the improved crnn network structure,a License Plate Sequence Recognition Networks(LPSR-Net)model is designed,which uses 312508 license plate images of synthetic data set and 115817 license plate images of CCPD 2019 data set,is tested on 45057 license plate images of CCPD 2019 data set,and its recognition accuracy reaches 99.7%,0.5 percentage points higher than the original CRNN model and 1.2 percentage points higher than the License Plate Recognition via Deep Neural Networks(LPRNet)model.(4)A variety of end-to-end license plate recognition algorithms are studied and compared with open source algorithms.Through experiments,it is found that the methods based on LPRetinaNet and LPSR-Net include positioning and recognition,and are offline at 1800 pixels wide by 640 pixels wide and 480 pixels high.The accuracy rate of the image data with license plate reached 96.05%,and the average time spent was only 81 milliseconds,16.33 percentage points higher than the HyperLPR open source license plate recognition system,and 16.05 percentage points higher than Lab-212-LPRS-v3.(5)By adding trigger module to capture the license plate image from the camera in real time,the license plate recognition system realized in this paper has achieved an average accuracy of 99.8% for the whole character and 407 milliseconds for the measured results of 1000 color images including license plate with the size of 1920 × 1080,which can basically meet the more robust and accurate vehicle in the campus scene Card identification.