车牌汉字识别技术的研究与实现

【作者】 朱峰；

【导师】 詹永照；

【作者基本信息】 江苏大学 ， 计算机应用技术， 2006， 硕士

【摘要】 车牌识别系统(LPR)是智能交通系统的核心组成部分，广泛应用于交通部门的违章检测(电子警察)、高速公路自动收费和智能停车场管理等方面。车牌识别系统主要包含车牌定位、字符分割、车牌字符识别三个主要部分，综合了模式识别、人工智能、计算机视觉和数字图形图像处理等多个学科领域。车牌汉字由于笔划多，相对于字母、数字来说，目前识别率较低，因此，车牌汉字识别问题是车牌识别的关键技术难题之一，研究车牌汉字识别问题并提高其识别率具有重要的现实意义。 本文针对车牌汉字识别及相关技术进行研究，主要工作包括： (1)在对车牌汉字进行识别前需要进行必要的预处理工作，我们首先运用车牌的先验知识进行粗定位，然后采用Otsu阈值二值化方法等进行快速鲁棒的车牌区域细定位，获得完整、清晰的车牌轮廓，并采用回扫式字符切割方法，充分利用了车牌字符固有的特点准确找到各个字符的分割点，为特征提取奠定了基础； (2)提出一种优化的Gabor滤波器特征提取算法直接对灰度图像进行特征提取。利用车牌汉字图像的统计信息和错误识别率进行参数的筛选解决了Gabor滤波器组用在车牌汉字特征提取中的参数优化设计问题，保证了较优的识别性能，也避免了传统方法中因二值化操作造成的结构信息丢失，又能有效降低数据维度，同时Gabor滤波器提取的特征对光照、轻微旋转具有很好的鲁棒性。 (3)提出采用改进BP神经网络作为分类器的分类方法，针对BP算法存在的收敛速度慢、易陷入局部极小的缺点，分别引入动量因子和自适应学习速率对其进行改进。该方法可以使网络收敛到更优点，同时加快了收敛速度。 (4)采用面向对象的设计方法给出了车牌汉字识别的原型系统，从实际运行实验中验证了上述方法的有效性。更多还原

【Abstract】 License plate Recognition(LPR) is the kernel resolution of Intelligent Transport System(ITS). It is widely used in the toll station,automatic payment highway fee,parking lot management,etc. LPR contains three sections — license plate localization,char extraction and char recognition. It integrated technologies such as pattern recognition, artificial intelligence, computer vision and image processing ,etc. This paper solves one of critical technology problems in LPR—license-plate Chinese character recognition, This study has biggish theory meaning and practice value.The main researches about Chinese character recognition of vehicle license plate and its corresponding technique in this paper are as follows:(1) The pre-processing for image is essential before character recognition. This paper firstly uses the experience knowledge to give a roughly orientation,then set a carefully orientation based on Otsu arithmetic,finally get the whole clear license plate.In order to solve the problem of character extraction,the fly-back method is put forward:the experience knowledge is made the best of.(2) An algorithm of feature extraction from grayscale image based on optimization Gabor filters is proposed. The performance of Gabor features depends strongly on filter parametes,in order to determine optimal values of filter parameters.we set parameters by the statistical information of Chinese character images and the leave-one-out method.The features extracted by this method which avoid losing the fabric of characters,are low-dimensional vectors.Compared with binary algorithm,this approach is robust and succeds.(3) Improving BP network has been applied into recognition system and much research about it has been done in this paper.BP network has slow convergence speed and is easy to fall into local least spot .To solve the two main problem, momentum gene and a self-adaptive learning rate have been imposed.Experiment result shows this algorithm can improve the recognition nicety of network.(4) At last,we design a prototype system of Chinese character on vehicle license plate recognition system based on object-orient methods and prove the effectivity of above algorithms of some experiments.更多还原