Research On Vehicle License Plate Detection And Recognition System Based On Deep Learning

In recent years,with the increasing number of cars and the continuous improvement of intelligent transportation scenarios,the importance of the technology of recognizing license plates in daily life is obvious.The traditional vehicle license plate recognition algorithms are based on three stages: license plate location recognition,license plate segmentation and license plate character recognition.It can be used effectively for ordinary scenes in daily life,but it is difficult to show strong robustness in the face of complex scenes such as image distortion and blur,so it is often unable to recognize the phenomenon.In recent years,more and more scholars have adopted deep learning methods to improve the accuracy of the license plate recognition.The small model,high precision,and embedded license plate recognition system has become the mainstream research direction.This paper uses deep learning models such as YOLOv5,U2-net and LPRnet to implement an effective license plate recognition system.This system is composed of two cascaded modules,which are license plate detection and character recognition systems.The former takes locating the license plate in the image and cutting the detection frame as the main task.The application of the Focus structure and CSP structure in the model backbone network significantly improves the accuracy of license plate positioning.The use of CIOU＿LOSS in the prediction terminal improves the detection accuracy and shortens the detection time.The performance of target detection model is tested by generating adversarial sample blocks and adding them to the license plate dataset.The latter takes the identification of the license plate characters in the detection frame as the main task.The license plate is corrected by U2-net binarization and perspective transformation,and the application of CTC loss and cluster search significantly enhances the character recognition effect.Compared with traditional license plate recognition methods,the algorithm in this article has significantly demonstrated the advantages of convolutional neural networks.It can not only be applied to conventional places such as parking lots and highways,but also under adverse conditions such as insufficient light at night,blurred images,and tilt.There is still a relatively stable robustness,which reduces the influence of the external environment on the license plate recognition and makes up for the deficiencies of the original research.This paper takes grouping experiment as the main test method,and adopts the method of controlling variables.Taking recall rate,accuracy rate and overall accuracy rate of the system as the main evaluation indicators,the test set is divided into six groups for license plate detection and character recognition experiments.Experiment with the overall system combining the above two systems and compare it with the RPNet network system.The detection results show that the average accuracy and processing time of the system studied in this paper are 94.3% and 240 ms,respectively,and can well complete the recognition in complex scenes and make up for practical applications.