Research On License Plate Recognition In Complex Scenes Based On Deep Learning

The importance of license plate recognition technology in real life is self-evident.The traditional license plate recognition algorithm is based on license plate location,character segmentation and character recognition,which can be effectively applied to specific scenes,but it is difficult to show reliable performance for license plate images in complex scenes.In recent years,deep learning has become a research hotspot,and the license plate recognition algorithm based on neural network has become the main research direction of academia.In this paper,an effective license plate recognition system is implemented by using Faster R-CNN and LSTM models in deep learning.The system is divided into two cascading parts.The first part is the license plate detection system,whose main function is to locate the license plate in the image.The second part is the license plate character recognition system.In the license plate detection algorithm,the introduction of RPN network greatly improves the positioning accuracy of the model.In the character recognition algorithm,the LSTM network transforms the character recognition problem into the sequence labeling problem,and reached better performance.Compared with the traditional method of license plate recognition algorithm,this paper give full play to the advantages of the depth of neural network,the most similar scenarios are not limited to existing system required by the fixed occasions,such as a garage or a car park entrance,but can be extended to all kinds of complicated scene,in the face of less ideal shooting conditions,such as blurring,inadequate illumination,all showed high robustness and accuracy.Finally,this paper also makes a lot of collation,screening and classification of CCPD data set,and supplements the sample set through GAN network and field collection.Thus,a license plate data set based on a variety of shooting scenes or image features is established.The basic experimental method in this paper is the grouping experimental method of controlled variables.The license plate images to be tested are divided into five categories for recognition test.The judgment indexes of recognition effect are recall rate and accuracy rate.Experiments were carried out on the license plate detection subsystem and license plate character recognition subsystem,and finally the whole experiment was carried out on the system composed of the two,and the test index of processing time was added.Experimental results show that the system described in this paper can realize license plate recognition in a variety of complex scenes,and achieved an average recall rate of 88.78% and accuracy rate of 98.27%,and the average processing time is also controlled within 200 ms,which can basically meet the requirements of practical application.