基于深度学习的车牌识别系统

发布时间：2018-03-10 20:44

本文选题：车牌识别　切入点：端到端算法　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着城市化发展,车牌自动化识别技术的需求日渐凸显。虽然在高清静止的场景中车牌识别已经较为成熟,但在更广义的场景中,车牌识别问题依然面临挑战。在本文中,我们使用近年来较为热门的深度学习技术对广义场景下中文车牌识别进行探究。在工程上,为了全面研究车牌问题,控制数据和标注的一致性,我们创建了一个车牌数据库系统。为了及时查看算法结果以及方便地进行修改标注,我们搭建了测试网站以及标注网站。在算法上,我们先调研了传统的车牌识别流程以及算法技巧。我们利用调研得到的方法以及深度学习方法来实现传统的车牌流程,即车牌检测、车牌定位、字符分割和字符识别。我们分析传统流程的缺点,发现瓶颈在于过多的流程和各个算法模块间衔接不够紧密积累起了速度以及准确率的严重损失。所以本文把车牌识别分为了车牌检测和正车牌识别两个部分,并提出了端到端的方法来克服传统流程进行正车牌识别带来的问题。为了支持深度学习端到端的正车牌识别模型的训练,本文又设计和实现了模拟车牌生成器来产生大量正车牌图片。本文搜集了大量车牌数据,并通过实验证明了在正车牌上,端到端的方法在速度以及准确率上均大幅度地优于传统流程方法。除此之外,在车牌检测上,本文比较了目前较为常用的两种物体检测方法,在车牌图像清晰化上,本文也进行了一些探索性的实验。
[Abstract]:With the development of urbanization, the demand for automatic license plate recognition technology is becoming increasingly prominent. Although license plate recognition is mature in the high-definition still scene, but in the broader scenario, the license plate recognition problem still faces challenges. In this paper, We use the popular deep learning technology in recent years to explore the Chinese license plate recognition in the generalized scene. In engineering, in order to study the license plate problem comprehensively, we control the consistency of the data and the label. We created a license plate database system. In order to see the results of the algorithm and easily modify the tagging, we built the testing website and tagging website. We first investigate the traditional license plate recognition process and algorithm skills. We use the method and depth learning method to achieve the traditional license plate process, that is, license plate detection, license plate location, Character segmentation and character recognition. We analyze the shortcomings of traditional processes, Finding the bottleneck lies in the excessive flow and the insufficient connection between the algorithm modules, which accumulates the serious loss of speed and accuracy. Therefore, this paper divides the license plate recognition into two parts: license plate detection and positive license plate recognition. An end-to-end approach is proposed to overcome the problems caused by the traditional process of positive license plate recognition. This paper also designs and implements a simulated license plate generator to generate a large number of positive license plate images. The end-to-end method is much better than the traditional method in speed and accuracy. In addition, in the license plate detection, this paper compares the two commonly used object detection methods, in the license plate image clarity, Some exploratory experiments are also carried out in this paper.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】