【作者】 俞伟广；

【导师】 赵建民；

【作者基本信息】 浙江师范大学 ， 计算机软件与理论， 2007， 硕士

【摘要】 车辆牌照识别是智能交通系统(ITS)的一个重要组成部分,尤其是复杂背景下的车牌识别,有着广泛的应用领域和美好的应用前景,其发展必将大大加速ITS进程。本文针对目前车牌识别算法的局限性,主要研究了复杂背景下的车牌识别技术。本文主要从以下几个方面进行研究:1)复杂背景下的车体检测和定位。本文采用瞬时块差分法来检测车体,通过对视频序列中连续两帧图像做以点为中心的3×3框架的块差分,并取差分的平方以加强运动信息,然后利用滤波、膨胀等技术准确定位车体。实验表明该算法具有很好的检测效果。2)车牌定位。先对车辆图像进行二值化并进行一系列数学形态学操作,再利用车牌先验信息进行过滤,产生车牌候选区域。然后对候选车牌区域进行彩色边缘检测,利用车牌区域固定的颜色对信息和结构信息准确定位车牌。算法在满足实时性的同时拥有极高的定位准确率。3)车牌字符分割。由于车牌区域除了汉字之外,其它的字符都是连通的,本文采用基于纹理信息和连通区域相结合的方法来提取单个字符。算法实现简单,而且可以在一定程度上消除预处理效果较差带来的影响,同时又可从根本上解决了倾斜车牌分割的问题。4)车牌字符识别。提出了基于Null Space PCA和改进的BP神经网络相结合的字符识别方法。首先对预处理过的字符图像进行Null Space PCA变换以达到降维的目的,然后把不同位的车牌字符分别放入汉字、字母、字母数字和数字四个不同的神经网络分类器,用改进的BP算法进行训练识别。实验结果表明该算法具有较高的识别率和识别速度,能够满足实时识别的要求。本文结合理论和实践,对复杂背景下的车牌识别系统的各关键技术进行了深入研究,并提出了新的思想和方法,对车牌字符识别技术的研究具有很大的意义和应用价值。更多还原

【Abstract】 License Plate Recognition is one important constituent part of the Intelligence Transportation System (ITS). Especially it is the license plate recognition in complicated contexts. It covers comprehensive application areas and enjoys a great application perspective. Its development is sure to accelerate the progress of ITS. In light of the weak points in contemporary license plate recognition algorithm, the present study focuses on the techniques of license plate recognition within complicated contexts.This thesis is divided into the following several aspects for the research: 1) the detection and location of the vehicle in complicated contexts. The present study employs the instantaneous difference method based on block, that is, to detect the vehicle by doing the block difference of the frame 3×3 with dots in the center for two successive images in the video array and then squaring the difference to strengthen the information of movement. The vehicle is accurately located with the help of techniques such as filter and dilation. The experiment has showed that this kind of algorithm holds sound testing effects. 2) license plate recognition. The first step is to make binary images of vehicles and to do a series of mathematic morphology operation. And then is to filter according to the existing information and to come up with the license plate areas for candidacy. After that, the color edge detection is made for the chosen license plate areas and license plates are accurately located according to fixed color-pair information and structural information in license plate areas. This kind of algorithm can reach the accurate location rate with a rather high degree as well as satisfy the real-time performance. 3) character segmentation of license plates. Because characters in the license plate areas are connected except for Chinese characters, the present study abstracts single characters by the way of combining information based on veins and connected domains. It is easy to do the algorithm, and to some extent, this can help eliminate the bad effect brought about by pretreatment as well as fundamentally solve the problem of slant license plates’ segmentation. 4) character recognition of license plates. This part puts forward a new way of character recognition which is based on Null Space PCA and the improved BP neural network. At first, for the aim of dimension, the pretreated character images should be made to undergo the transfer of Null Space PCA. Then different characters should be put respectively into four different classifiers of neural network (they are Chinese characters, letters, letters and numbers, and numbers). The result of the experiment shows that this kind of algorithm can reach the high recognition rate and recognition speed, which can satisfy the demand of recognition in the real-time performance.The present study combines both theories and practice, and delves into the various key techniques in the license plate recognition system under complicated contexts. New ideas and methods are given, which has great research significance and adds to the value of application for the character recognition technique of license plates.更多还原