车牌识别算法研究及系统设计

【作者】 张司兴；

【导师】 李跃华；

【作者基本信息】 南京理工大学 ， 电磁场与微波技术， 2010， 硕士

【摘要】 随着经济的发展,汽车数量急剧增加,智能交通系统(ITS)应运而生,车牌识别(LPR)系统是智能交通系统的重要组成部分,是目前模式识别研究领域的一个热点,具有很好的应用前景和发展潜力。车牌识别主要由图像采集、预处理、车牌定位、字符分割和字符识别几部分组成,本文从车牌定位、字符分割和字符识别三方面对车牌识别技术进行了详细的研究。论文的主要内容如下：1、介绍了车牌识别系统的应用背景、系统组成及发展现状。2、介绍了车牌图像的预处理方法,主要包括：彩色图像灰度化以及灰度变换、直方图均衡和中值滤波等图像增强方法。3、在介绍几种边缘检测算法的基础上,研究了一种基于边缘点区域连通的车牌定位算法,该算法首先求车辆图像的垂直边缘,进而将图像中的边缘点连接成若干个连通区域,然后利用车牌的几何特征与其他背景区域的区别,将非车牌区域剔除,从而完成车牌定位。4、在车牌定位完成后,对有倾斜的车牌进行倾斜矫正,车牌倾斜有两种倾斜：水平倾斜和垂直倾斜。对于水平倾斜矫正使用了Hough变换和K-L展开式两种矫正方法,并对这两种方法进行了对比。论文使用了基于剪切变换的方法来矫正垂直倾斜的车牌。倾斜矫正之后,去除车牌的上下边框,最后将车牌上的字符串分割成为单个的字符并将其归一化和细化,使字符的大小相同,字符笔划的宽度只有一个像素。5、在字符分割、归一化和细化的基础上,使用粗网格法对车牌字符进行特征提取,根据所选取的特征设计BP神经网络,并用该神经网络分类器进行字符识别。根据中国车牌上字符的排列特点,使用了两个分类器进行识别,第一个神经网络专门用来识别汉字,另外一个用来识别字母和数字,识别之后将两个分类器的输出结果组合就可以得到识别结果。6、介绍了车牌识别系统硬件平台TDSDM642EVM评估板,包括CPU、存储器、视频接口、视频编解码芯片和以太网口,使用该硬件平台,配合相应的外围硬件及软件构建一个车牌识别系统。7、在VC++平台上使用MFC构建了一个软件车牌识别系统,该系统能够对采集到的图像进行车牌识别,论文对其菜单及识别过程做了详细的演示。更多还原

【Abstract】 With the development of economy, the number of cars increased dramatically thus the Intelligent Transportation System(ITS) arises at the historic moment. License Plate Recognition(LPR) is an important part of ITS. The LPR System is a hotspot in Pattern Recognition research field and has good prospect and development potential.LPR System is combined by several parts such as image acquisition, image preprocessing, plate location、char split and char recognition. This paper studied the LPR in detail in plate location, char split and char recognition.The main content of the paper as follows:1、Introduced the application background, system configuration and current situation of the LPR System.2、Introduced several license plate image preprocessing methods such as color image graying and some image inhancement methods such as gray-scale transformation, histogram equalization and median filter.3、After the introduction of several edge detect algorithms, a plate location algorithm based on edge region connection was studied. This algorithm combines the edges of the image into some independent connected regions, then the areas which did not contain license plate were removed according to the geometrical characteristic between the licence plate and the background. Finally the position of the license plate was located accurately.4、After the location of license plate, it is necessary to rotate some license plates which were tilt. There are two kinds of tilt modes: one is horizontal tilt, the other one is vertical tilt. Methods based Hough transform and K-L expansion were used to correct horizontal tilt license plates and compare was did between these two methods. Then the vertical tilt licence plates were corrected by shear transformation. Then the frame of the license plate was removed. Finally the string on the plate was splited into single chars. After normalization and thining, these chars had the same size and the width of the strokes of the chars was one pixel.5、On the foundation of char spliting, coarse grid feature extraction was done on these chars. BP neural net was adopted to recognize these chars. According to the arrangement of the chars on the license plate, two classfiers one for Chinese chars and the other for combination of letters and numbers were used to recognize these chars. The recognition result was got after synthesizing the output of these two classfiers.6、Introduced the hardware platform TDSDM642EVM evaluation board which would be used to build the LPR system. CPU, memory, video ports,video codec and the ethernet were introduced. Coordinate with corresponding software and peripheral hardware, a LPR system could be built based on this board.7、Finally a LPR system was built on the VC++ platform by MFC tools. This system could recognize the license plates in the photos. The system menu and recognition procedure was shown in detail in the paper.更多还原