Research On License Plate Recognition Technology Based On Deep Learning

With the increasing demand of modern traffic technology for motor vehicle management,license plate recognition system is widely used in a variety of real scenes,such as the underground parking lot of a large supermarket,the entrance to the community,high-speed toll station.Due to the continuous optimization of the performance of the license plate recognition system,the examination and approval of vehicles in and out is accelerated,which greatly alleviates the problem of traffic congestion,saves human resources and improves work efficiency.At the same time,deep learning,as a new subset of machine learning,has emerged in speech recognition,computer vision,predictive recommendation and other fields.The combination of the two has certain research significance and practical value.The research content of this paper is composed of image preprocessing,license plate location,character segmentation and character recognition.The main work is as follows:1.Complex weather environment and Angle problems will make the image taken become fuzzy,which will have a great impact on the positioning and recognition results,while the image preprocessing part focuses on image blur and illumination inequality.In the process of contrastive analysis of the effect of relevant algorithms,this paper proposes the gray-scale and filtering de-noising operations according to the content requirements of this paper,and improves them with the gray-scale transformation on this basis,in order to enhance the contrast between the target license plate and the background information in the image.Finally,edge detection is used to finish the image processing;2.Analyze and compare the advantages and disadvantages of several common license plate location algorithms,and then propose a morphological region filling algorithm based on binary image according to the specifications and characteristics of domestic license plate.Among them,the license plate candidate region is extracted according to the rectangular characteristics of the license plate,and the obtained part is distinguished by fine positioning algorithm;3.Vertical projection algorithm is used to segment the extracted license plate region.This process is mainly based on the prior information of license plate,the continuous characteristics of characters,set the appropriate threshold to achieve the segmentation of characters;4.Based on the traditional neural network,this paper adopts the current mainstream deep learning algorithm for character recognition to meet the requirements of accuracy and robustness in the operation process.On the premise of fully considering the uniqueness of characters,the ReLu activation function is selected as the improvement of the network structure,so as to improve the recognition effect of license plate characters.