车牌定位与字符分割算法的研究与实现

【作者】 鲁飞；

【导师】 解梅；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2011， 硕士

【摘要】 进入21世纪以来,我国交通运输业得到了飞速的发展,并已成为国家大力发展的基础建设领域。这使得人们的生活越来越便捷,但是急剧增加的各种机动车的数量也给城市的交通管理和控制带来了极大的挑战。智能交通系统在这种背景下应运而生,成为当前交通管理发展的主要方向。其中,车牌识别系统是智能交通的关键技术,扮演着非常重要的角色。借助于车牌识别系统,我们便可以对车辆进行全天候的自动监控,这不仅提高了效率,而且节省了人力和物力。故可将其广泛应用于十字路口车辆管理系统、高速公路车辆管理系统、小区与停车场收费管理系统等领域当中。完整的车牌识别系统一般由软件和硬件构成。软件算法采用图像处理的相关技术分为三个部分:车牌定位算法,车牌字符分割算法,车牌字符识别算法。本文主要对其中的车牌定位算法,车牌字符分割算法进行研究探讨。1.车牌定位算法。车牌定位是进行车牌识别的第一步,是整个系统的基础。针对本文已有算法只能提取出蓝底白字的车牌且定位正确率不高的问题,本文从车牌图片自身的特点出发,提出了一种基于边缘纹理和边缘颜色对特征的车牌定位算法。该算法首先采用Sobel算子对车牌图像进行边缘检测,然后根据车牌边缘纹理特点和颜色对信息两次去除干扰边缘,最后通过滑窗遍历边缘图像获得车牌区域的连通域,从而成功从图片中定位出各种颜色搭配和单行或者双行排列的车牌。两次去除干扰边缘的操作,不仅简化了后续的处理,而且大大减少了伪车牌的出现。在对车牌进行倾斜校正时,针对已有的基于Hough变换的算法在车牌干扰严重的情况下校正失败的问题,本文对其进行了一些改进。另外,为了避免在精确定位车牌的左右边界时,边框的去除不理想造成字符分割失败的问题,本文动态设计出一个车牌大小的模板结合一阶差分计算得到较精确的车牌左右边界。2.车牌字符分割算法。在进行字符分割时,需要将车牌归一化为黑底白字。针对已有算法对字符笔画较饱满和字符之间存在粘连的车牌不能成功进行归一化的问题,本文提出了一种基于形态学的方法予以解决。为了进一步提高车牌字符分割正确率,本文尝试了一种基于支持向量机和车牌投影特征的车牌字符分割算法。该算法主要利用车牌的投影特征进行训练,分类。同时,在字符分割过程中还对定位结果中可能存在的伪车牌进行了去除。对于双行车牌,本文也一并完成了双行车牌的字符分割算法。本文所有的算法在MATLAB上进行了仿真实现。在对算法的测试中,以一个十字路口采集的大量车辆图片作为算法测试数据源对算法进行测试。测试结果表明车牌定位算法的准确率为95.1%,字符分割算法的准确率为95%。更多还原

【Abstract】 As the coming of the 21st century, the communication and transportation in our country have developed themselves rapidly and have stepped into the field of infrastructure, which the government would make great efforts to support. It makes our life become more and more convenient, but it also brings a huge challenge to the urban traffic because of more and more vehicles. Intelligent traffic emerges and becomes the main future development as the times require. As the critical technique, the license plate recognition (LPR) plays an important role in intelligent traffic. With the help of LPR, the vehicles can be supervised in all-weather automatically. It not only increases efficiency but also saves manpower and material resources. So it could be used widely in the occasions of intersection, expressway, parking place and so on.The whole license plate recognition system is composed of software system and hardware system. The software system can be divided into three parts according to the image processing technology: the license plate location, the segmentation of license plate characters and the recognition of characters. This article mainly discusses and researches on the algorithm of license plate location and segmentation of license plate characters.1. the algorithm of license plate location. This part is the first step and it plays a fundamental role in the whole system. The existing algorithm can only make the plate located, which has white lettering on a blue background. In view of the problem, this paper proposes a new locating method based on edge texture and edge-color pair. In this algorithm, we firstly obtain the edge of license plate with Sobel operator. Then remove the disturbed edge according to the character of edge texture and edge-color pair. At last, we get the connected region of candidate license plate by slip window and extract the LP. The operator of removing disturbed edge for double times not only makes the subsequent process easy but also minimizes the occurrence of pseudo plate greatly. In the process of precise location, the slant LP needs to be corrected. The existing method based on Hough transform often failed because of serious disturbance in LP. In our paper, we make some improvements to the existing algorithm. Lastly, in order to avoid the influence by imprecise left and borders of LP, which maybe make subsequent character segmentation failed. This paper combines template and first-order difference to get the accurate left and right borders of LP.2. The algorithm of segmentation of license plate characters. Before the character segmentation, we need to normalize the LP to state of black background with white text. Because of the adhesion between characters, the existing method will fail in the process of normalization. In our paper, the morphology method is employed to solve this problem. In order to improve the character segmentation accuracy, we try a new method based on support vector machine (SVM). This method mainly uses projection feature of LP to train and classify. In the process, we also further remove the pseudo plate. At the same time, we complete the segmentation of the double line LP.Our entire algorithm is fulfilled on the platform of MALAB. This paper uses a lot of images taken in the intersection to test the proposed approach. From the test result we know that the accuracy rate of license plate location is 95.1%, the accuracy of license plate segmentation is 95%.更多还原