License Plate Recognition System In Complex Scenes Based On Deep Learning Method

License plate recognition(LPR)is a comprehensive application of scene text recognition and target detection in the field of image processing.It plays a key role in intelligent criminal vehicle tracking system.At present,the commonly used license plate recognition system can be divided into four parts: LP detection,LP correction,character segmentation and character recognition.Almost of these methods can only be applied in simple specific scenes.And the distance and angle of the capture camera are fixed which leads to a great limitation in applications.Moreover,these methods with large time complexity are difficult to meet the application requirements of fast LP recognition in dynamic complex scenes.In order to realize a high-efficiency LPR system,we apply the deep learning technology in our recognition system to optimize the LPR process,which only requires three steps: LP detection,skew correction and character recognition.Firstly,the framework of single shot multi-box detector(SSD)is used to realize an end-to-end LP detection and classification network.Then,we propose a skew correction method based on multiple thresholding binarization.Finally,an end-to-end multi-character recognition is realized by using convolutional recurrent neural network(CRNN).Moreover,due to the large amount of parameters in deep learning network which leads to higher computation cost,this paper optimizes the calculations of the network layers and reorganize the internal structure of the SSD and CRNN and greatly improves the recognition efficiency.After experimenting on 1216 images with multiple scenes,the system achieves 96.13% recognition accuracy with 21 FPS average speed on PC platform and 8FPS on the Jetson TX1 embedded platform.The result shows that the proposed LPR system can be successfully applied in dynamic and complex scenes.Compared with traditional systems,the proposed LPR system has a wider practical application since the significant improvement of the robustness,accuracy and speed.