复杂背景下车牌识别算法的研究与实现

发布时间：2018-12-21 12:14

【摘要】：随着我国经济水平的不断上升,机动车保有量逐年提高,更多地服务于我们的生活。机动车辆的增多对道路使用、机动车管理提出了更高的要求。机动车号牌与车辆之间存在一对一的关系,因此通过车牌的识别可以更加有效地管理车辆。车牌识别技术研究近年来一直是智能交通系统研究的热门之一,但是目前市场上的车牌识别产品,其通用性不足,即在复杂环境如光照变化强烈、倾斜、污损、噪声等情况下车牌识别的效率不高。因此,研究复杂背景下的车牌识别技术有着重大的市场价值。基于上述背景,本文对车牌识别技术中的关键环节进行了深入的研究,主要包括车牌定位、倾斜校正、字符分割及字符识别等,结合现代计算机视觉的最新技术成果,对传统算法的缺点提出改进或提出新的算法,同时结合实验进行验证分析,取得了良好的效果。本文的主要内容如下:1、车牌定位算法。根据车牌区域颜色信息和边缘信息丰富的特点,算法先对边缘信息进行检测,再使用HSI空间的颜色分量对边缘信息进行筛选,接着使用自定义的边缘连接方法和形态学方法填补边缘信息,再使用连通域和车牌的几何特征选择车牌的候选区域。接着对候选区域进行精确定位处理,包括使用边缘信息进行倾斜校正、去除上下和左右边框。最后,使用SVM方法,对精确定位后的车牌提取HOG和LBP特征,进行伪车牌的剔除,从而定位出真的车牌区域。2、字符分割算法。本文对传统的两种算法提出了改进,详细论述了改进后的两种算法,即基于多阈值和连通域的字符分割算法、基于字符间距和二值投影的字符分割算法。前者针对连通域提取时无法一次性提取所有字符的特点,设计多阈值多次提取符合条件的字符区域,并提出了非连通汉字的提取方法。后者利用车牌字符间距的特点,找出车牌中第二个字符和第三个字符的分界点,然后使用二值投影和字符的几何特征进行字符分割。3、字符识别算法。引入卷积神经网络,针对汉字和字母数字的不同特征,设计不同的网络结构,从而对字符进行识别,同时该算法其鲁棒性好,提升了应对复杂环境的能力。本文的算法在VS2010平台上实现,编程语言C++,使用了计算机视觉工具库OpenCV 1.0。
[Abstract]:With the development of economy in our country, the quantity of motor vehicle has been increasing year by year and serving our life more. The increase of motor vehicles puts forward higher requirements for road use and vehicle management. There is a one-to-one relationship between vehicle license plate and vehicle, so vehicle can be managed more effectively by license plate recognition. In recent years, the research of license plate recognition technology has been one of the hot topics in the research of intelligent transportation system. However, at present, the license plate recognition products in the market are not universal enough, that is, in complex environments, such as intense changes of illumination, tilt, fouling, etc. The efficiency of license plate recognition is not high under the condition of noise and so on. Therefore, the research of license plate recognition technology in complex background has great market value. Based on the above background, this paper has carried on the thorough research to the license plate recognition technology key link, mainly includes the license plate localization, the tilt correction, the character segmentation and the character recognition and so on, unifies the modern computer vision newest technology achievement, An improvement or a new algorithm is proposed for the shortcomings of the traditional algorithm, and good results are obtained by combining the experimental results with the verification and analysis. The main contents of this paper are as follows: 1. License plate location algorithm. According to the rich color information and edge information of license plate region, the algorithm first detects the edge information, and then uses the color component of the HSI space to filter the edge information. Then the self-defined edge linking method and morphological method are used to fill the edge information, and then the connected domain and the geometric feature of the license plate are used to select the candidate region of the license plate. Then the candidate regions are accurately located, including edge information for tilt correction, removal of upper and lower edges and left and right borders. Finally, using the SVM method, the HOG and LBP features are extracted from the accurately located license plate, and the pseudo-license plate is removed, and then the true license plate region. 2. 2, character segmentation algorithm is obtained. In this paper, the two traditional algorithms are improved, and the two improved algorithms are discussed in detail, that is, the character segmentation algorithm based on multi-threshold and connected domain, the character segmentation algorithm based on character spacing and binary projection. In view of the feature that all characters can not be extracted at one time when the connected domain is extracted, the former designs multiple thresholds to extract the character regions that meet the requirements, and proposes a method for extracting disconnected Chinese characters. The latter uses the character spacing of license plate to find out the boundary point between the second character and the third character in the license plate, and then uses binary projection and the geometric feature of the character to segment the character. 3, character recognition algorithm. The convolution neural network is introduced to design different network structures for the different characteristics of Chinese characters and alphanumeric characters so as to recognize characters. At the same time the algorithm has good robustness and improves the ability to deal with complex environments. The algorithm of this paper is implemented on the VS2010 platform, programming language C, using the computer vision tool library OpenCV 1.0.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【参考文献】