基于图像处理的高速列车车号识别算法研究

发布时间：2018-06-25 21:49

本文选题：笔画宽度变换 + 图像处理　；参考：《西南交通大学》2017年硕士论文

【摘要】：目前,我国迎来了高速铁路建设的高潮,按照《中长期铁路网规划》我国最终形成"八横八纵"的高速铁路网,运营里程的增加和速度的提升使列车安全问题越来越受重视。受电弓滑板作为列车从接触网上获取电能的关键部件,在利用现有图像处理技术监测到其发生异常情况后,需要迅速判断列车车号等相关信息,及时排除安全隐患并上报有关部门。使用传统射频技术识别车号无法建立与受电弓滑板状态一一对应的关系,而利用摄像机同时采集车号图像与受电弓滑板图像进行车号识别达到了这样要求,是进一步检测受电弓故障异常状态的基础。本文工作是依托受电弓滑板监测装置来展开的,算法利用摄像机对列车侧面拍摄得到的车号数据,进行车号定位、矫正、分割后,完成了高速列车车号自动识别的任务。本文主要内容如下:(1)在图像预处理阶段,使用带彩色恢复的多尺度视网膜皮层增强算法(Multi-scale Retinex with Color Restoration,MSRCR)对数据集中部分受到天气或者拍摄环境影响产生的过曝、过暗以及存在雾气的图像进行图像增强。(2)车号精确定位过程中,针对笔画宽度变换(Stroke Width Transform,SWT)出现计算量大、定位不准等问题,首先提前进行图像轮廓筛选和设置最大笔画阈值,然后进行笔画宽度变换,最后采用形态学膨胀操作将单个字符笔画值连接为一个车号整体进行定位。经测试,改进后的SWT变换程序运算时间比原算法缩短了约70%。(3)针对车号存在透视畸变的问题,首先分析了传统基于灭点算法在本文中的不适应性,然后研究了变换不变低秩纹理(Transform Invariant Low-rank Textures,TILT)算法,建立了求解的数学模型并进行矫正,最后单个车号字符分割过程中,本文引入了一种基于字符间距的轮廓检测均分算法。该算法对存在断裂和粘连字符的分割正确率与普通的轮廓分割算法相比有较大的提高。(4)在单个字符识别过程中,首先对小样本数据进行数据增强以满足训练网络的需要,然后以经典LeNet_5网络结构为参考,设计了适合本文数据的卷积神经网络(Convolutional Neural Networks,CNN)网络结构,采用实验证明了网络的合理性。最终的实验结果表明,对11类字符分类准确率可以达到99.9%。
[Abstract]:At present, our country has ushered in the high tide of high-speed railway construction. According to the "medium and long term Railway Network Plan", our country finally formed the "eight horizontal and eight vertical" high-speed railway network. The increase of mileage and the increase of speed make the train safety problem more and more important. Pantograph slide plate is the key component of the train to obtain electric energy from the contact line. After using the existing image processing technology to monitor the abnormal situation, it is necessary to quickly judge the train number and other relevant information. Remove hidden safety troubles in time and report to relevant departments. Using the traditional radio frequency technology to identify the vehicle number can not establish a one-to-one relationship with the pantograph skateboard state, but using the camera to collect the vehicle number image and the pantograph skateboard image at the same time for the identification of the vehicle number can achieve this requirement. It is the basis of detecting the abnormal state of pantograph fault. The work of this paper is based on pantograph skateboard monitoring device. The algorithm uses video camera to obtain the vehicle number data from the side of the train to locate, correct and segment the vehicle number, and complete the task of automatic identification of the vehicle number of the high-speed train. The main contents of this paper are as follows: (1) Multi-scale Retinex with Color RestorationMSRCR (Multi-scale Retinex with Color RestorationMSRCRs) is used in the image preprocessing stage. (2) in the process of accurate location of vehicle number, the Stroke width transform SWT (Stroke width transform SWT) has the problems of large amount of calculation and inaccurate positioning. Firstly, the image contour is screened and the maximum stroke threshold is set in advance. Then the stroke width is transformed and the single character stroke value is connected to a vehicle number to locate by morphological expansion operation. After testing, the operation time of the improved SWT transform program is shortened by about 70 percent compared with the original algorithm. (3) aiming at the problem of the vehicle number with perspective distortion, this paper first analyzes the inadaptability of the traditional algorithm based on vanishing point in this paper. Then, the transform variant Low-rank texture (TILT) algorithm is studied, and the mathematical model is established and corrected. Finally, in the process of single vehicle number character segmentation, a contour detection averaging algorithm based on character spacing is introduced in this paper. Compared with the common contour segmentation algorithm, the segmentation accuracy of this algorithm is much higher than that of the common contour segmentation algorithm. (4) in the process of single character recognition, the small sample data is firstly enhanced to meet the needs of the training network. Then, a Convolutional Neural Networks (CNN) network structure suitable for the data of this paper is designed, and the rationality of the network is proved by experiments. The final experimental results show that the accuracy of classification of 11 kinds of characters can reach 99.9%.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】