【作者】 王润民；

【导师】 钱盛友；

【作者基本信息】 湖南师范大学 ， 物理电子学， 2007， 硕士

【摘要】 车牌自动识别(LPR)技术是是智能交通系统(ITS)中一项非常重要的技术。车牌识别系统主要包括三个部分：车牌定位、车牌字符分割和车牌字符识别。本文针对车牌识别系统的三个关键技术进行了研究并提出了相应的算法，论文研究工作具体体现在以下几个方面：(1)从国内车牌的特点出发，提出了一种基于能量滤波和小波的车牌定位方法。根据车牌在水平方向能量高且集中的特点构造一个能量函数，能量滤波后获取车牌的大致位置，再由小波分析和形态学方法准确确定车牌位置。仿真结果表明该方法取得了满意的效果。(2)针对车牌字符分割，提出了一种基于神经网络和颜色特征的车牌字符分割方法。该方法直接对车牌的彩色图像进行处理，在判别车牌类型的基础上，采用神经网络对车牌颜色进行识别，将彩色图像转化为二值图像，最后结合投影法和字符连通性特点对字符进行分割。与基于灰度图像的字符分割方法比较，该方法能更准确、清晰地分割字符。(3)在进行车牌字符识别时，特征向量的选取与维数对识别结果产生很大的影响。本文提出了一种基于小波包和Zernike矩特征提取的车牌字符识别方法。对小波包系数和重构后所得图像的Zernike矩所组成的特征空间进行降维处理后，将特征向量作为神经网络训练和分类的参数对车牌中的数字进行识别，实验结果表明选用本文特征向量识别效果良好。(4)根据国内汽车车牌中字符排列的特点，提出了一种基于SVM混合网络的车牌字符识别方法。首先构造了汉字识别子网、英文字母识别子网、英文字母与数字识别子网以及数字识别子网，并提取字符的小波包系数和Zernike矩做为特征向量，然后在各个子网中采用SVM方法对车牌字符进行识别。实验结果表明，采用本文方法的识别效果优于BP神经网络及RBF神经网络识别方法。(5)以高斯核为其核函数的支持向量机识别性能对惩罚因子C和核函数参数σ的选取是敏感的。针对高斯核支持向量机在车牌字符识别问题中的应用，提出了一种基于遗传算法的参数选择方法。利用遗传算法对支持向量机的参数进行优化，最后在各个子网中分别采用参数优化后的支持向量机对车牌字符进行识别，取得了令人满意的识别率。更多还原

【Abstract】 Vehicle license plate recognition technology is very important in Intelligent Transportation Systems. There are three primary parts in vehicle license plate recognition system: license plate location, license plate characters segmentation and license plate characters recognition. The three pivotal technologies are studied and corresponding settle methods are presented in this paper. The research work is developed in the following aspects.(1) A kind of location method of vehicle license plate based on energy filter and wavelet is presented in this paper. The energy function is constructed according to the high and concentrated energy in the horizontal direction. The vehicle license plate image is roughly located by energy filter, and the license plate is accurately located by wavelet analysis and a series of morphological operations. The experimental results demonstrate the efficiency of the proposed approach.(2) A kind of segmentation method of vehicle license plate characters based on neural networks and color feature is presented. The vehicle license plate images are binarized using BP neural network after the kinds of the vehicle license plates have been judged, the vehicle license plate characters are accurately segmented by using the projection method and the characters’ connexity. The experimental results demonstrate the efficiency of the proposed approach.(3) The kind and the dimension of the feature vectors have important infection on vehicle license plate recognition. A kind of character recognition method of the vehicle license plate based on the wavelet packet and zernike moments is presented in this paper. The wavelet packet coefficients and the zernike moments make up the feature space, which is processed by reducing the dimension. The digits of the vehicle license plate are recognized by BP neural network. The experimental results demonstrate the efficiency of the proposed approach.(4) A kind of character recognition method of the vehicle license plate based on support vector machines is presented in this paper. Firstly, a Chinese character recognition sub-network, a English character recognition sub-network, a Chinese character and English character recognition sub-network and a digital character recognition sub-network are constructed for Chinese vehicle license plate characters’ properties. Then, the characters are recognized by SVM in every sub-network. The experimental results: vehicle license plate character recognition using SVM is better than the BP neural networks and RBF neural networks.(5) The performance of SVM with Gauss kernel is influenced greatly by the penalty parameter C and the scale parameter o. A kind of method to select these parameters using genetic algorithms(GA) is proposed based on the study of vehicle license plate characters recognition. The characters of the vehicle license plate are recognized by SVM with optimized parameters in various sub-networks. The experimental results demonstrate the efficiency of the proposed approach.更多还原