论文总字数：20815字

摘 要

随着社会的迅速发展，私人交通工具的广泛普及，汽车在人们的生活中扮演着越来越不可缺少的角色，是居家旅行出远门的必备之物，极大地解放和发展了生产力。与此同时也对城市的交通管理提出了相当大的考验。车牌是车辆的唯一身份证明，是联系车辆与车主的重要桥梁。

车牌识别的研究事实上从上个世纪就已经开始，这是一个涉及图像处理、模式识别等领域的研究课题。利用计算机技术将摄像头拍摄到的车辆信息进行处理，或者对视频图像的分析，从而获得车辆的身份证--车牌号码。当前车牌识别的方法有很多种，图像处理方面有open CV（Open source Computer Vision Library,开放源代码计算机视觉库），matplotlib库等，而神经网络框架有tensorflow、Caffe、Torch等。

本文提出采用深度学习的方法来进行车牌识别。所谓深度学习就是指采用多层神经网络、模仿人类大脑新皮层中的神经元层的活动来进行数据处理的方法。CNN（Convolutional Neural Network）是常见的深度学习模型之一，其在图像识别领域（例如：人脸、交通标志的识别等）被证明是极其有效的方法。

本文的工作主要包含以下几个方面：

1. 车牌的检测、定位以及预处理：现实生活中所获得的包含车牌的图像往往是光线不均匀的、角度有倾斜的。首先利用训练好的分类器将车牌块从包含车牌的图像中框出来，然后计算车牌的倾斜角，对车牌进行校正。矫正完了以后就要对车牌进行一系列的预处理，包括灰度化、去噪、二值化等。
2. 字符分割分割：完成车牌检测、定位及预处理之后就要对黑白车牌块进行分割。一是将7个字符分割开；二是将分割开的字符图片分别保存。在预处理部分通过计算车牌倾斜角度再用霍夫变换等方法将车牌块校正，然后按照150*40的比例压缩好就 可以根据比例将每个字符切割下来分别保存。
3. 基于CNN进行识别：本文搜集处理了几万张固定大小与格式的字符图片，分成省份、城市代号、数字三类，三类图片要先按32*40的比例压缩，再进行灰度化、去噪、二值化分别输入到设计好参数的三个CNN中进行训练。训练好之后测试识别的准确率。

关键字： 车牌检测 深度学习 CNN

License Plate Recognition Based on Deep Learning

ABSTRACT

With the rapid development of society and the widespread popularity of private transport, automobiles play an increasingly indispensable role in people's lives. They are essential for home travel and travel, which greatly liberates and develops the productive forces. At the same time, they put forward a considerable test for urban traffic management. License plate is the only identity certificate of a vehicle and an important bridge between the vehicle and its owner.

In fact, the research of license plate recognition has been started since the last century, which is a research topic involving image processing, pattern recognition and other fields. Computer technology is used to process the vehicle information captured by the camera, or to analyze the video image, so as to obtain the vehicle ID card - license plate number. At present, there are many methods of license plate recognition, such as open CV, Matplotlib library and so on, while the framework of neural network includes tensorflow, Caffe, Torch and so on.

In this paper, a deep learning method is proposed for license plate recognition. The so-called in-depth learning refers to the method of data processing by using multi-layer neural network to imitate the activity of the neuron layer in the new cortex of human brain. CNN (Convolutional Neural Network) is one of the common deep learning models. It has been proved to be an extremely effective method in the field of image recognition (such as face recognition, traffic sign recognition, etc.).

The work of this paper mainly includes the following aspects:

1. License Plate Detection, Location and Preprocessing: In real life, the images containing license plates are often uneven in light and inclined in angle. Firstly, the trained classifier is used to frame the license plate from the image containing the license plate, and then the tilt angle of the license plate is calculated to correct the license plate. After the correction, a series of pretreatment of license plate, including grayscale, denoising, binarization and so on, will be carried out.

2. Character segmentation: After completing license plate detection, location and preprocessing, black and white license plate blocks are segmented. One is to divide seven characters; the other is to save separated character pictures separately. In the pretreatment part, the license plate is corrected by calculating the tilt angle of the license plate and Hough transform. Then, each character can be cut and saved separately according to the ratio of 136*36.

3. Recognition based on CNN: This paper collects and processes tens of thousands of character pictures of fixed size and format, which are divided into three categories: province, city code and number. The three types of pictures should be compressed by 32*40 ratio first, then gray, denoising and binarization are input into three CNNs of designed parameters for training. After training, the accuracy of recognition was tested.

keywords: license plate deep learning CNN

目录

摘要 I

ABSTRACT III

第一章 绪论 1

1.1研究背景和意义 1

1.2车牌识别的研究现状 2

1.3车牌识别的流程 4

1.4本文的主要内容和章节安排 5

第二章 神经网络及深度学习的基本理论 7

2.1 神经网络概述 7

2.2 深度学习 9

2.3 本章总结 13

第三章 车牌识别具体流程 14

3.1 汽车车牌的定位 14

3.2 字符分割 19

3.3字符识别 20

3.4本章总结 21

第四章 基于深度学习的车牌识别实现 22

4.1 相关第三方库介绍 22

4.2 本项目方案 23

4.3 实验设置 27

4.4 实验结果和分析 28

4.5 本章总结 30

第五章 总结与展望 31

参考文献 33

致谢 35

第一章 绪论

1.1研究背景和意义

城市化的迅速发展也导致着交通压力的倍增，如果只是采用人力去管理的化必然成本极高，而且伴随着第四次工业革命的到来，智慧交通必然是今后发展的大趋势。现代车牌识别技术利用最先进的图像处理技术并结合了人工智能方法已经可以做到实时提取车辆图片并且实时获取车辆信息。监督车辆行为、实时车流调控、掌握道路信息是智慧城市系统在Iot中在交通领域的主要应用,而智能交通系统第一要素就是对车辆身份的采集。车牌识别技术是智能交通中最重要的组成部分,在城市道路监控、电子不停车收费、小区出入控制等重要场合中发挥着举足轻重的作用，是实现智慧城市的重要手段之一[1]。因此，车牌识别的重要性不言而喻，是未来的智慧城市中进行身份认证的必要前提。在当今AI大热门的环境下，计算机视觉、图像处理成为重要的研究方向，且在神经网络的助力下有了巨大的进步，达到了商用的程度。

请支付后下载全文，论文总字数：20815字