License Plate Detection And Recognition Based On Deep Learning

There are a large number of cars in China and the number of vehicles increased year by year.This phenomenon greatly facilitates people's trips and brings convenience to people's lives,but it also brings a series of problems,such as running red lights,illegal parking,speeding overload and traffic jams.These problems not only lead to traffic chaos,but also cause major traffic safety accidents which will endanger people's lives.In order to solve the above problems more efficiently,we must vigorously develop intelligent transportation,and strengthen supervision and management through intelligent traffic,and discover and solve problems in a timely manner.A very important part of the intelligent transportation system is the license plate detection and recognition.In this paper,the deep learning method is used to study the license plate detection and recognition.The traditional license plate detection methods are mostly using the sliding window method.This method is not efficient,and the effect is poor.In this paper,we compare the methods of target detection of deep learning in recent years.These methods include R-CNN,SPP-Net,Fast-R-CNN,Faster-R-CNN,YOLO and SSD.R-CNN,SPP-Net,Fast-R-CNN and Faster-R-CNN are proposal methods.These methods need to extract proposal regions and then classify them and transform proposal regions' location to final license plate location.YOLO and SSD are regression-based methods that do not require explicit extraction of proposal regions,so the speed is much faster.YOLO and SSD are both regression-based method,but YOLO only uses a single feature map for full connection operation,which makes YOLO's recall rate relatively low.SSD uses small convolution in the multi-scale feature map,which makes the SSD's recall rate and accuracy relatively high.Therefore,this article uses SSD to detect the license plate,and compares it with the Faster-R-CNN's result.As we all know,it is hard to collect a large number of license plate,so using a synthetic method to generate a large number of license plate samples in this paper.In this paper,the license plate synthesis method can be described as this:first,filling in the blank license plate template with the license plate characters,and then simulating the license plate shooting situation.The simulation methods include license plate blur,brightness and contrast adjustment,perspective transformation and Gaussian noise.Finally,CNN license plate recognition model is trained using a large number of synthetic license plate samples,and the license plate recognition performance is tested on the test set.In the license plate detection experiment,two SSD models,SSD300 and SSD512,were trained and tested on three test datasets.The AP @ 0.5,AP @ 0.6,AP @ 0.7,AP@ 0.8 and FPS are evaluation index.The FPS of the SSD300 is between 46 and 51,while the FPS of the Faster-R-CNN is between 3 and 5.At most IoU thresholds,the AP of the SSD300 is better than the Faster-R-CNN.The AP of the SSD512 model is the highest among the three models,and it's FPS is between 26 and 29.In addition,as the IoU threshold continues to increase,Faster-R-CNN's AP drops rapidly,while the SSD model is slower.In the license plate recognition experiment,the license plate recognition model which is trained in this paper has reached the correct recognition rate of 99%or more in both data set A and data set B,and the FPS is between 71 and 72.