基于人工神经网络的车牌 识别

H.Erdinc Kocer（埃尔丁克·科塞尔）^一个， K.库尔萨特·切维克^b

^一个塞尔丘克大学技术教育学院， 科尼亚 42250， 土耳其

^bNigde Uni. Bor Vocational High School， Nigde 51700， 土耳其

抽象

近年来，由于 交通中的车辆数量不断增加，个人在交通管制中工作的必要性 正在增加。 为了解决这个问题 ，正在 开发基于计算机的自动控制系统。 其中一个系统是自动车牌识别系统。本文提出了一种基于人工神经网络的自动车牌识别系统。在这个系统中，使用了259张车辆图片。这些车辆照片是从CCD相机拍摄的，然后使用 图像处理算法从该图片中确定尺寸为220x50 像素 的车牌区域。 采用Canny边缘检测算子和斑点着色法对车牌中的字母、数字等字符进行定位和确定。 将斑点着色方法应用于 ROI以分离 字符。在这项工作的最后阶段，使用平均绝对偏差公式提取了字符特征。然后使用前馈传播的多层感知器神经网络对数字化字符进行分类。上一节给出了正确的分类率。

⃝c 2010 由 Elsevier Ltd. 出版，在 CC BY-NC-ND 许可下开放获取。

由客座编辑负责 的选择和/或同行评审。

关键词：车牌 识别、人工神经网络、斑点着色、字符识别

介绍

无花果。 1. 自动车牌识别系统的块方案

这项工作 的 布局可以分为 7 个部分。在第 1 节中， 介绍了有关这项工作 的一般信息， 之前关于车牌识别的工作是 如第 2 节中的表格所示 。 在第 3 节中，本地化

1877-0509 ⃝c 2010 由 Elsevier Ltd. 出版，在 CC BY-NC-ND 许可下开放获取。

doi：10.1016/j.procs.2010.12.169

描述了过程。第 4 节介绍了字符的分割过程。在第5节中，介绍了分段字符的特征提取过程。第 6 节介绍了使用 ANN 识别字符的方法。实验结果在本书的最后一节中介绍。

过往作品

之前的作品将根据土耳其民用车牌识别的作品举行。板区定位（PRL）、字符分割（CS）和字符识别（CR）过程的成功率（SR）见表1[1-6]。

表 1. 以往作品的成功率

板块区域的定位

车牌识别系统的第一阶段是从车辆图像中查找车牌位置。板区域通常由白色背景和黑色字符组成。因此，该区域的黑白颜色之间的过渡非常密集。找到包含大多数过渡点的区域就足以定位板块区域。

为此，将Canny边缘检测算子应用于车辆图像以获取过渡点。Canny 边缘检测器使用基于高斯平滑一阶导数的滤波器。在平滑图像并消除噪点后，下一步是通过获取图像的渐变来找到边缘强度。对于此过程，此运算符使用 3x3 维度的矩阵。 然后计算梯度 的边缘 强度。 这些信息为我们提供了边缘点，因此可以确定密集过渡点区域。然后从此边缘图中确定黑色和白色之间的 过渡点。 图2a、图2b和图2c分别显示了原始 、检测到的边缘和局部板区域图像。

无花果。 2. （a） 汽车原件; （b） 检测到边缘; （c） 局部板块区域

字符的 分割

在分割过程之前，应增强灰度级车牌图像。因为在通过相机拍摄图像时可能会出现对比度差异。此外，不需要的脏污区域可能会被放置在板上，这些噪声会向负方向影响分割过程。

在这项工作中，通过应用对比度扩展和中值滤波技术增强了灰度级板图像。因此，图像之间的对比度差异和 噪点之间的差异，例如白色背景中的脏区域

板可以 消除。 在图像增强阶段之后， 实现斑点着色方法 确定字符的边界。

Contrast extension
对比度扩展

To extend the contrast of an image means equalization of the histogram of that image. In other words, the contrast extension makes the image sharpen. The gray-level histogram of an image is the distribution of the gray level values in an image. The histogram equalization is a popular technique to improve the appearance of a poor contrasted image.
扩展图像的对比度意味着该图像的直方图均衡。换句话说，对比度扩展使图像更加清晰。图像的灰度直方图是图像中灰度值的分布。直方图均衡是一种流行的技术，用于改善对比度差的图像的外观。

The process of equalizing the histogram of an image consists of 4 steps [7]: (1) Find the sum of the histogram values. (2) Normalize these values dividing by the total number of pixels. (3) Multiply these normalized values by the maximum gray-level value. (4) Map the new gray level values. The contrast extended license plate image is shown in Fig 3b.
均衡图像直方图的过程包括4个步骤[7]：（1）求直方图值的总和。（2） 将这些值除以像素总数归一化。（3） 将这些归一化值乘以最大灰度值。（4） 映射新的灰度级别值。对比扩展车牌图像如图3b所示。

Median filtering
中值过滤

Median filter is used for eliminating the unwanted noisy regions. In this filtering method, the 3x3 matrices is passed around the image. The dimension of this matrices can be adjusted according to the noise level.
中值滤波器用于消除不需要的噪声区域。在这种过滤方法中，3x3 矩阵在图像周围传递。该矩阵的尺寸可以根据噪声水平进行调整。

The process is working as [7]; (1) one pixel is chosen as center pixel of the 3x3 matrices, (2) the arounding pixels are assigned as neigborhood pixels, (3) the sorting process are employed between these nine pixels from smaller to the bigger, (4) the fifth element is assigned as median element, (5) these procedures are implemented to the all pixels in plate image. The filtered plate image is shown in Fig 3c.
该过程的工作方式为 [7];（1）选择一个像素作为3x3矩阵的中心像素，（2）将周围的像素指定为邻接像素，（3）在这九个像素之间采用从小到大的排序过程，（4）将第五个元素指定为中间元素，（5）将这些程序应用于板图像中的所有像素。滤波后的板图像如图3c所示。

Fig. 3. (a) original plate region image; (b) contrast extended image; (c) median filtered image
图3. （a）原始板区图像;（b） 扩展图像;（c） 过滤图像的中值

The blob coloring method
斑点着色方法

The blob (Binary Large Object) coloring algorithm has a strong architecture to determine the closed and contacless regions in a binary image. This algorithm uses a special L shaped template to scan the image from left to right and from up to down. This scanning process determine the independent regions by obtaining the connections into four direction from zero valued background. In this work, four directional blob coloring algorithm is applied to the binary coding license plate image for getting the characters [8]. After implementation, the segmented characters were obtained from the license plate region image (Fig 4).
blob（二进制大对象）着色算法具有强大的体系结构，用于确定二进制图像中的闭合区域和无触觉区域。该算法使用特殊的 L 形模板从左到右、从上到下扫描图像。该扫描过程通过从零值背景获取四个方向的连接来确定独立区域。本文采用四向斑点着色算法对二进制编码车牌图像进行分类[8]。实现后，从车牌区域图像中获取分段字符（图4）。

Fig. 4. The segmented characters
图 4.分段的字符

In this work, the segmented characters were classified as numbers and letters separately. For this purpose, the plate image was divided into three region. The first region consists of two digit numbers that indicates the city traffic code. The second region consists of one-to-three digit letters. The third region consists of two-to-four digit numbers. The plate image was scanned form left to right horizontally and the spaces between characters were determined in this process. If the value of the space is higher than threshold value than the character region is signed. The numbers were localized as 28x35 pixels dimension. The letters were localized as 30x40 pixels dimension. Some samples of numbers and letters segmented from plate region are shown in Fig 5.
在这项工作中，分段的字符分别被分类为数字和字母。为此，将板图像分为三个区域。第一个区域由两位数字组成，表示城市交通代码。第二个区域由一到三位数字的字母组成。第三个区域由两到四位数字组成。从左到右水平扫描印版图像，并在此过程中确定字符之间的间距。如果空格的值高于阈值，则对字符区域进行签名。这些数字被本地化为 28x35 像素尺寸。这些字母被本地化为 30x40 像素尺寸。从板区域分割出来的一些数字和字母样本如图5所示。

Fig. 5. Some samples of the segmented characters
图 5.分段字符的一些示例

Feature Extraction
特征提取

In this study, the obtained characters were saved as an image file separately. The dimension of the numbers was determined as 28x35 pixels, the dimension of the letters was determined as 30x40 pixels. The numbers and the letters were classified by using two separate ANN for increasing the success rate of the recognition phase. Before classification, the character images should be feature extracted. Feature extraction provides us to obtain the most discriminating information of an image. This information can be presented as a feature vector. A feature vector that includes global and local features of an character should be encoded so that the comparison between characters can be made. In the proposed approach, the feature vector of an iris image was encoded by using Average Absolute Deviation algorithm [9]. This algorithm is defined as:
在这项研究中，将获得的字符单独保存为图像文件。数字的尺寸确定为 28x35 像素，字母的尺寸确定为 30x40 像素。使用两个单独的人工神经网络对数字和字母进行分类，以提高识别阶段的成功率。在分类之前，应提取字符图像的特征。特征提取为我们提供了获得图像最有鉴别力的信息。此信息可以呈现为特征向量。应对包含字符的全局和局部特征的特征向量进行编码，以便可以对字符进行比较。在所提出的方法中，使用平均绝对偏差算法对虹膜图像的特征向量进行编码[9]。此算法定义为：

V  1  f (x, y)  m 

(1)

 N 

where N is the number of pixels in the image, m is the mean of the image and f(x,y) is the value at point (x,y). In this work, the number images were divided into 4x5 pixels dimensioned sub-images and the letter images were divided into 5x5 pixels dimensioned sub-images. Each sub-image were feature extracted by applying AAD. We obtained the feature vectors with the length of 49 byte for numbers and 48 byte for letters. The entire feature vectors were applied to the ANN as an input for classification of the characters.
其中 N 是图像中的像素数，m 是图像的平均值，f（x，y） 是点 （x，y） 处的值。在这项工作中，将数字图像划分为4x5像素维度的子图像，将字母图像划分为5x5像素维度的子图像。每个子图像都是通过应用 AAD 提取的特征。我们得到了数字长度为 49 字节、字母长度为 48 字节的特征向量。将整个特征向量应用于 ANN，作为字符分类的输入。

Recognition of the Characters
字符的识别

In our work, the numbers and the letters were classified by using two separate ANN for increasing the success rate of the recognition phase. Both of them have same architecture but only the input numbers were differed. The reason for using two separate ANN for recognition is preventing the complexity of recognition of similar numbers and letters such as “0” – “O”, “2” – “Z” and “8” – “B”. As we can know, this complexity will decrease the recognition success.
在我们的工作中，使用两个单独的人工神经网络对数字和字母进行分类，以提高识别阶段的成功率。它们都具有相同的架构，但只有输入数字不同。使用两个单独的人工神经网络进行识别的原因是防止识别相似数字和字母的复杂性，例如“0” – “O”， “2” – “Z” 和 “8” – “B”。正如我们所知，这种复杂性会降低识别成功率。

In the proposed approach, a multi layered perceptron (MLP) ANN model was used for classification of the characters. The processing units in MLP are arranged in three layers. These are input layer (includes the information you would use to make decision), hidden layer (helps network to compute more complicated associations) and output layer (includes the resulting decision) [10,11]. Each neuron in the input layer is fed directly to the hidden layer neurons via a series of weights. The sum of the products of the weights and the inputs is calculated in each node. The calculated values are fed directly to the output layer neurons via a series of weights. As in hidden layer, the sum of the products of the weights and the hidden layer neuron outputs is calculated in each node in the output layer. If the error between calculated output value and the desired value is more than the error ratio, then the training (changing the weights and calculating the new output by using the new weights) process begins. This training process can be finished by obtaining the desired error rate for all input combinations.
在所提出的方法中，使用多层感知器（MLP）ANN模型对字符进行分类。MLP中的处理单元分为三层。它们是输入层（包括用于决策的信息）、隐藏层（帮助网络计算更复杂的关联）和输出层（包括最终决策）[10,11]。输入层中的每个神经元通过一系列权重直接馈送到隐藏层神经元。在每个节点中计算权重和输入的乘积之和。计算出的值通过一系列权重直接馈送到输出层神经元。与隐藏层一样，在输出层的每个节点中计算权重和隐藏层神经元输出的乘积之和。如果计算出的输出值与所需值之间的误差大于误差比，则开始训练（更改权重并使用新权重计算新输出）过程。可以通过获得所有输入组合所需的错误率来完成此训练过程。

For training the ANN, feed-forward back-propagation algorithm was chosen. For measuring the training performance of the network, mean square error (MSE) function is used. The value of the MSE is used to determine how well the network output fits the desired output. The stop criteria for supervised training are usually based on MSE. Most often the training is set to terminate when the MSE drops to some threshold. Approaching the MSE value to the zero means that the calculated output value is becoming the closer to the desired output value.
为了训练ANN，选择了前馈反向传播算法。为了测量网络的训练性能，使用了均方误差 （MSE） 函数。MSE 的值用于确定网络输出与所需输出的拟合程度。监督培训的停止标准通常基于 MSE。大多数情况下，训练设置为在 MSE 下降到某个阈值时终止。将 MSE 值接近零意味着计算出的输出值越来越接近所需的输出值。

Experimental Results
实验结果

In order to evaluate the performance of the proposed system, 259 vehicle images were employed. Sigmoid
为了评估所提出的系统的性能，使用了259张车辆图像。乙状结肠

function is used in the activation of neurons. Quick back propagation learning algorithm was used for training the ANN. Maximum 5000 iterations were performed for each input set. When the system reach to the minimum error rate which defined by the user, the iterations will be stopped. The defined minimum error rate for this application is 0,001. Only one of the input images was used for testing the system, the rest was performed in training phase. The iteration-MSE graphics of best results for each character data set are shown in Fig 6. The training reaches the minimum error rate in 4457 iterations for the numbers and 1180 iterations for the letters.
功能用于神经元的激活。采用快速反向传播学习算法对ANN进行训练，每个输入集最多进行5000次迭代。当系统达到用户定义的最小错误率时，迭代将停止。此应用程序定义的最小错误率为 0,001。只有一个输入图像用于测试系统，其余的在训练阶段执行。图 6 显示了每个字符数据集的最佳结果的迭代 MSE 图形。训练在数字的 4457 次迭代和字母的 1180 次迭代中达到最小错误率。

Fig. 6. MSE – Iterations graphics of the training process
图 6.MSE – 训练过程的迭代图形

The success rates for the plate region localization (PRL), character segmentation (CS) and character recognition (CR) stages of the proposed system are given in Table 2. As a result, 247 license plates in 259 vehicle image were recognized correctly in this work, so the overall recognition percentage of the system is 95,36%.
表2给出了所提系统的板区定位（PRL）、字符分割（CS）和字符识别（CR）阶段的成功率。结果，在这项工作中，247车辆图像中的259个车牌被正确识别，因此该系统的总体识别率为95,36%。

Table 2. The success rates of the proposed automatic license plate recognition system.
表 2.所提出的自动车牌识别系统的成功率。

References
引用

H. Caner, “FPGA Donanimi Uzerinde Arac Plakasi Tanima Sistemi”, Ms. Thesis, Pages 58-65, Hacettepe Uni., 2006.
H. Caner，“FPGA 硬件上的车牌识别系统”，硕士论文，第 58-65 页，Hacettepe Uni.，2006 年。

S. Ozbay, “Automatic Vehicle Identification by Plate Recognition”, Ms. Thesis, Pages 79-80, GaziantepUni., 2006.
S. Ozbay，“通过车牌识别自动识别车辆”，女士论文，第79-80页，加济安泰普大学，2006年。

G. Yavuz, “Plaka Tanima Sistemi”, Ms. Thesis, Pages 59-60, Sakarya Uni., 2008.
G. Yavuz，“Plaka Tanima Sistemi”，女士论文，第59-60页，萨卡里亚大学，2008年。

B. Yalim, “Turk Sivil Plaka Standartlari Icin Arac Plaka Tanima Sistemi”, Ms. Thesis, Pages 88-91, Gazi Uni., 2008.
B. Yalim，“Turk sivil plaka standartlari icin arac plaka tanima sistemi”，硕士论文，第88-91页，Gazi Uni.，2008年。

I. Irmakci, “Otomatik Arac Plaka Tanima Sistemi”, Ms. Thesis, Pages 75-76, Ege Uni., 2008.
I. Irmakci，“Otomatik Arac Plaka Tanima Sistemi”，女士论文，第75-76页，Ege Uni.，2008年。

K. Bora, “Car Plate Recognition”, Ms. Thesis, Pages 40-41, Atilim Uni., 2009.
K. Bora，“车牌识别”，硕士论文，第 40-41 页，Atilim Uni.，2009 年。

S. Umbaugh, Computer vision and Image processing, Chapter 4, Prentice Hall, New Jersey, 1999.
S. Umbaugh，计算机视觉和图像处理，第 4 章，新泽西州 Prentice Hall，1999 年。

D.H. Ballard, C. M. Brown, Computer Vision, Pages 151-152, Prentice Hall, New Jersey, 1982.
D.H. Ballard， C. M. Brown， Computer Vision， Pages 151-152， Prentice Hall， New Jersey， 1982.

L. Ma, Y. Wang, T. Tan, “Iris recognition based on Multichannel Gabor Filtering”, The 5th Asian Conference on Computer Vision, Australia, 2002.
L. 马， Y. Wang， T. Tan，“基于多通道Gabor滤波的虹膜识别”，第五届亚洲计算机视觉会议，澳大利亚，2002年。

C.M.Bishop, Neural Networks for Pattern Recognition, Oxford Uni. Press, England, 1995.
C.M.Bishop，《用于模式识别的神经网络》，牛津大学出版社，英国，1995年。

Y. Ozbay, B. Karlik, “A fast training back propagation algorithm on windows”, Proceedings of 3rd Int. Symp. Mathematical & Computational Applications, Turkey, 2002, pp 204-210.
Y. Ozbay， B. Karlik， “A fast training back propagation algorithm on windows”， Proceedings of 3rd Int. Symp. Mathematical & Computational Applications， 土耳其， 2002， pp 204-210.