【作者】 张秀丽；

【导师】 杨英；

【作者基本信息】 东北大学 ， 车辆工程， 2008， 硕士

【摘要】 车牌识别(License Plate Recognition, LPR)系统是智能交通的重要组成部分,主要应用在电子计费领域,例如高速公路不停车收费、停车场管理、多用途收费系统。虽然系统已发展多年,但由于光照变化范围大、图像背景复杂等原因,仍有许多技术难点未能解决。本文对车牌照识别算法中的车牌定位、字符分割和字符识别三个方面的内容进行研究,设计并实现了相应的算法,以期解决系统的技术难点,主要研究成果与创新如下：本文对于彩色图像灰度化,图像增强,中值滤波等图像处理方法在车牌图像识别技术中的应用特点,进行了分析和总结。本文从车牌几何形状的特征出发,设计出一种首先基于颜色特征和车牌的先验知识对车牌进行粗定位,然后根据车牌边界的特点寻找目标对象特征点,对车牌进行精确定位。即在粗定位的基础之上,结合直方图特征分析方法对图像进行二值化,再采用旋转投影法寻找车牌的倾斜角度,进行倾斜校正,边框切除,最后跟据边界黑白像素跳变特点精确定位车牌。这样,通过以上的方法,可以满足本文对车牌定位的处理要求。对校正后的图像采用回扫式分割以及垂直投影方法,利用车牌几何特征从投影图中寻找各个字符的位置以实现字符分割。针对汉字中的二分字和三分字分割时,易错分割的问题,采用回扫式分割方法。并考虑到字符粘连,利用车牌单个字符尺寸的特征,硬性分割；针对字符左右偏移,采用修正车牌的实际宽度的方法；并将其与模板匹配分割等方法进行比较,得出本文的分割方法取得较好的分割结果。研究了基于模板匹配的统计模式识别的方法,以及基于泰勒公式的多模板建模的模式识别的方法。分析了两种方法的优缺点,并根据具体情况给出了改进方案。针对光照不均匀的情况,采用基于模板匹配的统计模式识别和基于泰勒公式的多模板建模的模式识别相结合的方法,实现了车牌字符识别。当光线均匀时,采用基于传统的模板匹配的识别方法；当光线不均匀时,采用基于泰勒公式的多模板识别的识别方法。其中,泰勒公式多模板建模识别根据鲁棒性回归对特征误差最小化原则,得到相应的恢复系数向量；对权值欧式距离进行了分析,找出距离最小量,得出识别结果。同时本文对容易混淆的字符二次识别,使车牌识别识别率有较好的适应性。更多还原

【Abstract】 License plate recognition (PLR) system is an important part of intelligent transportation system, which is mostly applied to electronic payment system, such as non-stop toll collection in highway, non-attended parking fee payment, and multi-use payment. LPR system has been developed for many years. The two factors of the complicated conditions and various illuminations that affect the quality of scene images are the techniques puzzles to be solved. The three models of license plate located, characters segmented, and characters recognition are researched, then the relative algorithm is proposed in order to solve the relative techniques. Research achievements and innovations are as follow:The colorful image which is changed into the gray one, image enhancement, median-filter and so on, are analyzed and summarized in the plate image recognition application characters. Searching for the features of the object is designed based on the features of gray arrangement. The plate is located rough by the color feature and the plate information according to the geometry feature of the plates. Then the plate is located accurately by the way of searching for the object feature points according to the boundary of the white and black pixels. Then the binarization of the image is obtained by the histogram feature analysis method. In addition, the incline angle is found by the rotation projection method, then the incline of the image is adjusted; Moreover, the rim of the image is wiped off in order to satisfy the requirement of the plate location.The flyback segmentation and vertical projection are used to deal with the revised image. The characters are isolated according to the geometry feature of the plate to find the position of the single character. Next, the single character measure feature is used to isolate the conglutination of the characters. The adjusted actual plate width is applied aiming at the incline characters. The flyback segmentation is used to isolate the Chinese characters made up of the two elements or three elements. Then, the method of templates matching is compared with the method told above. The method in this article acquires a good segmentation according to the experiments.Analysing the pattern recognition of multipliable templates modeling based on the Taylor formula(MTMT) and the statistical pattern recognition based on traditional templates matching(TTM), the improvement is projected according to the factual condition. An algorithm combined the pattern recognition of multipliable templates modeling based on the Taylor formula(MTMT) and the statistical pattern recognition based on traditional templates matching(TTM) is used to recognize the characters, aiming at the nonuniform illumination. Then, the pattern recognition of multipliable templates modeling based on the Taylor formula(MTMT) uses the robust regression to minimize the feature errors, and the recognition result is obtained by analyzing the reconstructed weights errors. Moreover, the characters that are easy to recognize wrong is recognized secondly by TTM. In addition, the characters recognized wrong and the typical characters are able to add to the swatch in order to raise the recognition rate.更多还原