【作者】 黄翔星；

【导师】 王方石；

【作者基本信息】 北京交通大学 ， 计算机科学与技术， 2010， 硕士

【摘要】 车牌识别是智能交通系统中最关键的研究课题,有着广泛的应用前景,如交通道路监控、高速公路自动收费、停车场管理等。随着经济社会发展,机动车辆日益增加,对车辆进行安全管理、交通流引导和控制的需求越来越明显,因此研究更为稳定、快速、有效的车牌识别系统具有巨大的社会意义和实用价值。车牌识别关键技术由三部分组成：车牌定位、字符分割和字符识别。本文综合应用图像处理技术、模式识别、人工神经网络等方法,对这三大技术进行深入学习与研究,提出了有效的改进方法,并利用VC++6.0平台,编程实现了车牌识别系统。具体包括以下内容：在综合分析了各种典型车牌定位算法后,提出一种基于水平垂直投影的车牌定位方法。该方法对预处理后的图像进行一次平滑处理去除噪声,然后阈值化,再对阈值化后的二值图像在水平垂直方向上进行投影运算,求得车牌区域上下左右边缘位置后,进行裁剪。该方法特征提取简单、计算量小、速度快,且易于理解操作,能有效地实现车牌定位。在综合分析了基于投影、聚类、模板匹配等字符分割算法后,提出一种基于车牌先验知识的行列扫描字符分割方法。该方法结合我国车牌长宽、字符特征,先利用扫描行法获得车牌字符上下边界,再利用扫描列法依次获得各字符左右边界,以此分割各字符。实验证明,该方法执行速度快,能很好地处理由于车牌磨损、污染造成的字符粘连现象。在综合分析了基于模板匹配、人工神经网络的字符识别方法,以及深入研究了传统BP神经网络缺陷后,提出一种改进了的分组BP神经网络字符识别方法。该方法结合现有车牌字符类型,将BP神经网络分成四个子网络进行学习识别,并在权值修改时加入动量系数,很好地解决了学习速率过大或过小所引起的网络易发散或收敛慢的问题。经大量实验证明,该方法能有效提高BP网络学习速度以及字符识别准确率。更多还原

【Abstract】 License Plate Recognition is one of the most critical research topics in the Intelligent Transportation System. It has a broad application prospects, such as road traffic monitoring, automated highway toll collection, parking management, and so on. With the development of economic and society, the quantity of vehicles is increasing, and the demand of vehicle safety management, traffic guidance and control is more and more obviously. Therefore, it has great social significance and pratical value to make a research in more stable, fast and effective license plate recognition system.License plate recognition technology consists of three key components, i.e. license plate location, character segmentation and character recognition. This article used image processing and artificial neural network technologies, proposed effective improvements, and used VC++6.0 platform programming of the license plat recognition system. It included contents as follows:A new algorithm of license plate location based on horizontal and vertical projection is proposed. Firstly, after pre-processing, the image was smoothed to remove noise. Secondly, the image was projected in the horizontal and vertical direction after thresholding of binary image, when we got the upper and lower edge of the license plate area location, we could intercept the image. This method is simple, fast, and easy to understand, it can effectively locate the license plate.This article proposed a new method of license plate character segmentation which based on prior knowledge of plate. The method useed the fixed features of our license plate. Firstly, the upper and lower boundary of license characters was obtained by scanning each line of the plate. Secondly, the left and right boundary of license characters was obtained by scanning each column. Experiment results show that this method performs fast, and can effectively deal with the uncleared characters that caused by wear and pollution phenomena.Based on the analysis of template matching and artificial neural network character recognition methods. The article proposed a method of improved packet BP neural network character recognition. The method divided the BP neural network into four sub-networks, and added the momentum factor when weight changing. In the traditional BP network, when learning rate is too big or small, the network is easy to divergence or slow convergence. The packet BP network can slove this problem. Experiment results demonstrate that the method can effectively improve the speed of BP network training and character recognition accuracy.更多还原