基于卷积神经网络的智能车牌识别系统研究外文翻译资料

基于卷积神经网络的智能车牌识别系统研究

摘要：由于过去几十年人口的快速增长和人类对交通工具使用需求的增加。考虑通过牌照来识别这些车辆。随着图像处理技术的快速发展，通过采集车辆的图像来识别车牌，可以很好地解决这一问题。大多数交通应用都依赖于车牌的自动检测和识别技术，如搜索被盗车辆、道路交通管制、监控进入区域的车辆、车辆的一些信息、停车系统、监控边界路口、最高车速或闯红灯罚单，以及识别驾驶员的身份等。本文试图设计一种基于图像识别的新型车牌字符识别系统，实现对新型车牌字符的识别。 首先使用数码相机捕获车辆图片，并使用光学字符识别技术识别捕获图片的文本，然后将其与数据库中存储的所有车牌号码进行比较。

关键词：车牌识别；字符识别；神经网络；概率神经网络；图像处理

1.介绍

车牌识别在交通流量控制、自动停车系统、自动过桥系统、基于雷达的速度控制等领域有着广泛的应用。车牌识别系统的优点是它能够在不需要在汽车上安装额外设备的情况下操作。

牌照识别系统基本上具有三个模块，用于：

a）使用汽车的图像来定位车牌区域，

B）字符的提取，以及

C）使用合适的算法识别字符。

在这项工作中，开发了一种新的算法来识别不同光照、距离和倾斜条件下的车牌。使用Otsu的阈值化方法来确定给定汽车图像中可能找到牌照的区域。然后利用车牌特征检测出真实车牌区域。随后，执行列和向量（CSV）计算以实现字符提取。使用概率神经网络（PNN）来识别这些字符。在仿真中使用了MATLAB程序。

本节的其余部分总结了文献中以前的一些工作。

El-Adawi，Keshk和Haragi设计了一个基于神经网络的自动车牌识别系统，该系统使用反向传播算法进行训练。他们获得了89%的车牌提取成功率和93%的车牌字符识别成功率[1]。

Emiris和Koulouriotis将字母数字字符的训练和识别结合起来，用于半结构化环境中。他们列出了各种条件下字母A识别的成功率，平均成功率在59-77.4%之间[2]。

Juntanasub和Surreerattanan研究了两级泰国板块系统。他们采用离线Hausdorff距离技术进行相似性度量，从而进行识别。他们在车牌识别中获得了92%的成功率[3]。

Rattanathammawat和Chalidabhongse研究了板块定位问题。他们采用了Sobel边缘检测器、移动窗口板定位器和时间分析器来抑制错误检测。正确检测板的成功率为94%[4]。

Raus和Kreft在检测和字符识别中都使用了神经网络。他们将他们的方法与经典方法进行了比较，发现他们的方法更优越[5]。

Sirithinaphong和Chamnongthai利用机动车法规来训练他们的四层反向传播神经网络。车牌定位成功率为84.29%，字符识别成功率为80%[6]。

Park，Kim，Jung和Kim已经使用神经网络进行板块定位。他们在两组数据上测试了他们的网络，定位成功率分别为97.5%和99%[7]。

Kim，Kim，Kim和Kim通过神经网络分割车牌，并通过支持向量机识别字符。它们的分割率为97.5%，字符识别率为97.2%[8]。

剑锋、少发、志斌等曾从事中国车牌系统研究。他们采用神经网络进行颜色分析，以正确提取图版。其成功率为95.7%[9]。

Broumandia和Fathy已经使用神经网络来识别波斯语板块。他们在车牌识别中获得了平均95%的成功率[10]。

Ganapathy和Lui在字符识别中使用了前馈反向传播神经网络。其成功率为95%[11]。

2.车牌识别的概率神经网络方法

该算法适用于灰度图像。首先，将所讨论的图像转换为其灰度级。然后对灰度图像进行预处理。它们是底帽滤波、Otsu阈值、打开、标记、关闭、歪斜校正和可能包含板的矩形区域的提取。然后，使用板特征来确定正确的板区域。列和向量（CSV）图表用于字符分割。最后，利用概率神经网络对字符进行识别。该算法的流程图如图所示。1.

使用底帽滤波来增强潜在的板区域。阈值化用于灰度图像的二值化，从而将感兴趣的对象从背景中分离出来。由于环境因素，亮度水平可能会发生变化，因此需要进行一些调整。使用Otsu的阈值技术是因为其自适应特性。二值图像的片段根据它们的颜色被标记以实现分类。通过计算CSV及其局部最小值来完成板提取。应当将局部最小值与阈值进行比较。它们用于确定字母数字字符。

由字符提取模块区分的字符被隔离并保存在存储器中。对字符的尺寸进行均衡，并使用模板匹配算法。计算所提取的每个字符与数据库中的每个模板的相关性。对分割后的字符进行平滑处理，并计算边界处的临界

3.模拟结果

用于这项工作的数据库由260张汽车照片组成。每张照片的尺寸都是固定的，为384*288毫米。这些照片是在一天的不同时间，从不同的距离和不同的角度拍摄的。取自数据库的一些示例如图所示。

Fig. 2. Samples from the car image database.

对图像进行灰度化、滤波和阈值化等预处理。为了消除彩色图像中不必要的信息，进行了灰度级的还原。通过这种方式，处理速度大大提高。图6中示出了一个例子。

Fig. 3. (a) original picture; (b) gray level image

板块国产化是第一步，也是最重要的一步。如果它失败了，其余的肯定会失败。基于诸如纵横比和像素数量的特征来评估潜在的板区域，然后识别正确的板区域。在特征提取步骤中，使用列和向量来确定字符的边界。已经开发了一种算法来分离两个相邻字符，并将拆分为两个字符的字符统一起来。最后，利用概率神经网络对所获得的字符进行识别。从数据库中的图像中获得的特征被用作训练集，并且该训练集被用于训练神经网络。

将Bottom-hat滤波器应用于灰度图像，以增强潜在的车牌区域。在Otsu阈值之后，应用打开来删除不属于字符的小区域。然后贴标签和关闭是连续应用的。最后，纠正歪斜。对于来自数据库的样本图像，每个步骤的结果图像如图所示。

Fig.4 (a)gray level image; (b) Bottom-Hat filter;(c) Otsursquo;s thresholding; (d) opening; (e) labeling; (f) closing; (g) skew correction

如果获得多个候选板区域，则选择基于板特征。这些是：

1.印版高度应至少为12像素，

2.板宽应至少为16像素，

3.板的高度应不超过图像的1/8，

4.板宽应不超过图像的1/3，

5.板材面积应不超过总面积的1/4。

示例中获得的板材区域、相应的CSV图表和提取的字符如所示图5。字符位于CSV图表中两个局部极小值之间。分段字符如图所示。6.板区域和分割字符的另一个示例如图所示

Fig. 5. (a) extracted plate region; (b) corresponding CSV chart; (c) segmented characters

Fig. 6. Segmented characters ofthe example

Fig. 7. Second example (a) extracted plate region; (b) segmented characters

提取的特征被输入到概率神经网络中。输出是识别的字符。

4.结论

该程序在Intelreg;Coretrade;2 Duo处理器CPU P8400（2.26GHz，2267MHz）计算机上平均需要0.1秒来识别每个板。在车牌识别和字符识别的基础上，对仿真结果进行了评价。计算结果总结在表1中。

即使单个字符出错，也无法正确识别车牌。因此，字符识别是车牌识别的核心。字符识别模块的成功率的增加可以进一步提高车牌识别的成功率。今后，将继续沿着这一方向进行研究。

外文原文资料信息

1. 外文原文作者：INSODE 2011

2.外文原文所在书名或论文题目：A New License Plate Recognition System Based on Probabilistic Neural Networks

3.外文原文来源：2212-0173 copy; 2012 Published by Elsevier Ltd. Open access under CC BY-NC-ND license. doi:10.1016/j.protcy.2012.02.024

网页地址：https://www.sciencedirect.com/science/article/pii/S2212017312000254

References

1. Mohamed El-Adawi, Hesham Abd el Moneim Keshk and Mona Mahmoud Haragi, Automatic License Plate Recognition, IEEE Transactions on Intelligent Transport Systems, Vol. 5 (2004) 42-53.

2. D. M. Emiris and D. E. Koulouriotis, Automated Optic Recognition of Alphanumeric Content in Car License Plates in a Semi-Structured Environment, Proceedings of the International Conference on Image Processing, Vol. 3 (2001) 50–53.

3. R. Juntanasub and N. Sureerattanan, Car License Plate Recognition through Hausdorff Distance Technique, Proceedings of the IEEE ICTAIrsquo;05 (2005) 607-612.

4. P. Rattanathammawat and T. H. Chalidabhongse, A Car Plate D

剩余内容已隐藏，支付完成后下载完整资料

Available online at www.sciencedirect.com

Procedia Technology 1 (2012) 124 – 128

INSODE 2011

A New License Plate Recognition System Based on Probabilistic Neural Networks

Fikriye Ouml;ztuuml;rka, Figen Ouml;zena*

aHaliccedil; University, Electrical and Electronics Engineering Department, Şişli, Istanbul, Turkey

Abstract

A license plate recognition system employs image processing techniques, to help to identify the vehicles through their plates. License plate recognition is a process, where first the license plate region is localized in a car image supplied by one camera or by multiple cameras, and then the characters on the plate are identified by a character recognition system. There are many applications of the license plate recognition systems, both public and private. The algorithms, hardware and the network structure for recognition are designed according to the specific application. Recently, thanks to the advances in science and technology, the algorithms and hardware of higher quality have been designed, and license plate recognition systems are now widely used. The recognition can be done in three major steps: Localization of the plate, extraction of the plate characters, and recognition of the characters using a suitable identification method. In this paper, an algorithm is designed that can recognize plates using the pictures taken at various angles, various distances and different times of the day, thus under various illumination conditions. The plate is localized using Otsursquo;s thresholding method and the plate features. Vertical and horizontal histograms are used for character segmentation. Finally, character recognition is done by Probabilistic Neural Networks. Simulation results are included and performance analyses are tabulated. MATLAB program is used in the simulations.

Keywords: License plate recognition; character recognition; neural networks; probabilistic neural networks; image processing

1. Introduction

License plate recognition has many applications such as in traffic flow control, automatic parking systems, automatic bridge systems, and radar based speed control. The advantage ofthe license plate recognition system is its ability to operate without the need to install extra equipment on the car.

A license plate recognition system has basically three modules for:

a) Localization ofthe plate region using the image ofthe car,

b) Extraction ofthe characters, and

c) Recognition ofthe characters using a suitable algorithm.

In this work, a new algorithm has been developed to recognize the plates under varying illumination, distance and skew conditions. The regions of a given car image, where the license plate might be found, are determined using Otsus thresholding method. Then the plate features are used to detect the real plate region. Later, column sum vector (CSV) calculations are performed to implement the character extraction. Those characters are recognized using a Probabilistic Neural Network (PNN). MATLAB program is used in the simulations.

* Fikriye Ouml;ztuuml;rkTel.: 90 212 3430872; fax: 90 212 3430878.

E-mail address: figenozen@halic.edu.tr.

2212-0173 copy; 2012 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.

doi:10.1016/j.protcy.2012.02.024

Fikriye Ouml;ztuuml;rk and Figen Ouml;zenn /Procedia Technology 1 (2012) 124 – 128 125

Some ofthe previous work in the literature is summarized in the rest ofthis section.

El-Adawi, Keshk and Haragi have designed an automatic license plate recognition system based on Neural Networks that were trained using back propagation algorithm. They have obtained 89% success rate for license plate extraction and 93% success rate for character recognition of the extracted plates [1].

Emiris and Koulouriotis have combined training and recognition of alphanumeric characters to use in a semi - structured environment. They have tabulated the success rates for the recognition of letter A under various conditions, the averages for the success rates turned out to in the range of 59-77.4% [2].

Juntanasub and Surreerattanan have worked on two-level Thai plate system. They have employed off-line Hausdorff Distance technique for similarity measurement, and thus for recognition. They have obtained 92% success rate in plate recognition [3].

Rattanathammawat and Chalidabhongse have worked on plate localization problem. They have employed Sobel edge detector, moving-window plate localizer and a temporal analyzer for false detection suppression. The success rate for correct detection of plates was 94% [4].

Raus and Kreft have utilized Neural Networks, both in detection and character recognition. They have compared their method with the classical approaches and found out that theirs was superior [5].

Sirithinaphong and Chamnongthai have utilized Motor Vehicle Regulation to train their four-layer back propagation Neural Network. They have obtained 84.29% success rate for the plate localization and 80% success rate for the character recognition [6].

Park, Kim, Jung and Kim have used Neural Networks for plate localization. They have tested their networks on two sets of data and the success rates for localization were 97.5 and 99% [7]

Kim, Kim, Kim and Kim segmented the plates by Neural Networks and recognized the characters by Support Vector Machines. Their segmentation rate w

剩余内容已隐藏，支付完成后下载完整资料

资料编号：[595946]，资料为PDF文档或Word文档，PDF文档可免费转换为Word