Application Of Deep Learning Algorithm In License Plate Recognition System

LPRS(License plate recognition systems) is widely used in real life, such as toll stations, parking lot and intersections. These specific applications can alleviate phenomenon of the traffic congestion effectively, and also save the labor costs, improve efficiency and management role. At the same time, LPRS is an important part of the intelligent transportation system. Deep learning algorithm is the hotspot in the field of machine learning, which is widely used in image recognition, speech recognition, face recognition, natural language processing and so on. Therefore, this study has theoretical significance and practical value to a certain degree.This system consists of the license plate localization, character segmentation and character recognition. Based on the study of several common algorithms of each module, an improved algorithm with deep learning is purposed and a LPRS is achieved. The main work of the thesis is as follows:1. Comparing and analyzing the structure feature among deep belief network, stack auto-encoder network and convolution neural network, their advantages and disadvantages in image process field are pointed out, then establish the convolutional neural network as the application in this paper.2. Based on the study of the specifications of domestic license plate and several common localization algorithms, an improved method is proposed. According to the design principle of localization algorithm in this paper, which is the use of digital image technology in image suspected license plate region extraction work quickly to avoid deep learning algorithms taking full image scanning. Furthermore, the extraction candidate region is based on two significant features of the license plate(edges and color), then roughly select it combined with rectangular characteristic. For the candidate license region, Alex Net model performed well in classification is utilized to decide and retain the true license plate.3. As a result of deep learning is not fit to character segmentation algorithm, the author achieves the character segmentation algorithm which combined with the collected license, and based on several existing character segmentation algorithms. The algorithm uses a conventional digital image technology. That is to say, first, by using Hough transform, OTSU threshold segmentation and Radon transform, it make character segment easier, then implement license plate character segmentation based on prior information and projection features.4. According to the stringent requirements for robustness and recognition performance, deep learning is used to improve the recognition algorithm in this paper. with understanding of the traditional Le Net-5 model, author make several changes and have trained two networks to recognize the Numbers/Letters and Chinese characters respectively.5. LPRS is designed with hardware part and software part. Then using the Open CV library and other technology in the VS development environment, author realizes the algorithms proposed in this paper and successfully applied to the system.