License Plate Detection And Recognition Based On Deep Learning

With the rapid development of social science and technology,many cities have put forward the requirements of building smart cities.Smart transportation comes into being with the proposal of smart cities.The traditional license plate area detection system has many problems,such as large model,high computation,difficult deployment and low recognition rate in different scenes.These problems have led to the traditional license plate detection and recognition system can not meet the requirements of the new era.In this thesis,license plate recognition is carried out through two steps of license plate region detection and license plate character recognition.The improved lightweight target detection model is used to improve the accuracy and speed of license plate region detection,and the end-to-end text recognition combined model is used to accurately identify license plate characters.The specific contents of the research on street parking recognition in this thesis are as follows:1.Aiming at the problems of slow speed and low accuracy of license plate area detection,a license plate area detection strategy based on improved YOLOV4-tiny is proposed.Firstly,the Cross Stage Partial(Cross Stage Partial,CSP)module in YOLOV4-tiny network was replaced by Auxiliary Res Block(Auxiliary Res Block,Au-Res Block)module to reduce the model parameters for the convenience of embedded deployment and reduction of model size.Secondly,an information enhancement module was added to enhance feature information extraction by channel and spatial attention mechanism,and an appropriate loss function was selected to accelerate model convergence.Finally,a small size detection scale is added to the original network detection scale to improve the detection accuracy.In the experiment part,for the problem of scarce data sets,the algorithm model is trained by using the self-made data set of road cameras.Through the analysis of the experimental data,the proposed license plate region location detection and recognition model achieves the best balance between accuracy and speed.2.To solve the problem of low recognition rate of license plate character recognition,a license plate recognition strategy based on Light-Dense Net-GRU is proposed.In order to avoid the complex work of character segmentation,an end-to-end license plate character recognition algorithm was adopted.Based on the mainstream convolutional cyclic neural network,a Light-Dense Net algorithm was proposed to solve the problems of the convolutional layer algorithm with too many parameters,large model and imperfect feature extraction.In this method,Dense Net network is utilized to achieve feature reuse,and Dense blocks in the network are improved and lightweight modules are added to enrich the information of feature extraction and reduce model parameters.Aiming at the low recognition rate of tilted license plate characters,a spatial transformation network is added to correct the license plate to improve the recognition accuracy.In the cyclic layer,Gated Recurrent Unit(Gated Recurrent Unit,GRU)was used to predict character series instead of short and long-term memory network.Through the analysis of experimental data,the detection speed and accuracy of this recognition model are better than those of the mainstream models in various complex scenes.3.Through overall design,module design,system test and effect analysis,the algorithm of the previous chapters is integrated,and GUI design is carried out to make the recognition result more intuitive.At the same time,different scenes are tested in the system test part,and experimental results show that the recognition accuracy reaches 94.7%.Finally,compared with the traditional character segmentation license plate recognition system,the recognition rate is improved by 2.4%,and the recognition time is improved by 58.26 ms.Both the accuracy and speed of license plate recognition are superior.