基于CNN的载货列车信息识别系统设计与实现

发布时间：2018-06-12 17:55

本文选题：文字识别 + 图像处理　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：为了方便提高铁路货运管理的工作效率,减少企业对货运列车管理的投入。本文中利用文字识别相关技术对货运站点的火车车厢信息进行抓拍识别,并对识别记录进行存储管理。该方式改变了以往轨道衡值班人员在户外条件下对车厢载重、自重和容积等属性信息进行人工的记录并手动录入计算机存储的现状。同时减少了人工作业记录车厢信息出现的误差。本文中实现的系统可以有效地记录并管理车厢信息,大大降低人为因素的干预,同时减轻了轨道衡值班人员的工作量,节省企业对此项工作的人力投入。本文中利用网络摄像机、光电传感器等设备实现了一套完整的针对货运列车信息的识别系统。通过利用货运列车行进过程中车厢间隙的特征,结合一对光电光感器研究实现了一种针对货车车厢文字的控制抓拍方法。利用现场架设的多部摄像机对车厢两侧的文字信息进行抓拍,通过将车厢两侧不同质量的文字图像识别结果进行对比,以此来提高文字识别效率,其中识别的结果包括火车车型、车厢号、载重、自重、容积、宽高、换长等信息。系统同时利用射频识别(Radio Frequency Identification,RFID)通信技术对识别结果进行补充完善。结合现有成熟的视频监控手段,在传感器或识别功能出现故障时对过衡时的录像进行慢镜头回放,由值班人员根据系统录像回放补充车号、载重等信息。在文字识别部分中,对铁路货运列车车厢文字信息的识别进行研究,由于该应用场景的文字具有笔画不连续、笔画间隔大且受环境因素腐蚀严重等特点,利用传统的模版匹配或几何特征抽取等方法不能达到很好的识别效果,本文是选择卷积神经网络(Convolutional Neural Networks,CNN),通过前期图像分割得到大量数据样本进行训练识别。其中图像分割处理流程则是对原始图像利用以铃木算法为核心进行轮廓提取后确定文字区域。在根据文字区域的边缘信息水平投影,结合文字固定的宽高比例得到遍历模版后完成单个文字图像分割。通过系统现场实际测试得出,本文中设计的系统及采用的识别方法可以快速准确的识别指定场景中的文字信息,货车信息识别效果能达到应用标准。
[Abstract]:In order to improve the efficiency of railway freight management and reduce the investment of freight train management. In this paper, the relevant technology of character recognition is used to capture the train compartment information of freight station, and to store and manage the identification record. This method changed the situation of manual recording and manual storage of the load, weight and volume of the carriage under outdoor conditions. At the same time, the error of recording carriage information by manual operation is reduced. The system realized in this paper can effectively record and manage the information of the carriage, greatly reduce the intervention of human factors, at the same time reduce the workload of the personnel on duty of the track scale, and save the manpower input of the enterprise in this work. In this paper, a complete identification system for freight train information is realized by using network camera, photoelectric sensor and other equipment. Based on the characteristics of the gap between the freight trains and a pair of optoelectronic light sensors, a method of controlling and capturing the characters of the freight cars is presented in this paper. The text information on both sides of the car was captured by using a number of cameras set up on the spot, and the result of text image recognition of different quality on both sides of the car was compared to improve the efficiency of character recognition. The results include train model, car number, load, weight, volume, width and length. At the same time, the system uses radio frequency identification / radio frequency identification (RFID) communication technology to supplement and improve the identification results. Combined with the existing mature means of video surveillance, the slow motion video when the sensor or recognition function fails is played back in slow motion, and the personnel on duty play back the information of vehicle number and load according to the video recording of the system. In the part of character recognition, the text information recognition of railway freight train carriage is studied. The characters of the application scene are characterized by discontinuous strokes, large stroke intervals and serious corrosion by environmental factors. Traditional methods such as template matching or geometric feature extraction can not achieve a good recognition effect. In this paper, Convolutional Neural Networks CNNs are selected and a large number of data samples are obtained by image segmentation for training and recognition. The process of image segmentation is to use Suzuki algorithm as the core to extract the contour of the original image and determine the text area. Based on the horizontal projection of the edge information of the text region and the fixed width and height ratio of the text, the traversal template is obtained, and the segmentation of a single text image is completed. Through the field test of the system, it is concluded that the system designed in this paper and the recognition method used in this paper can quickly and accurately recognize the text information in the specified scene, and the recognition effect of the truck information can reach the application standard.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U29-39;TP391.41

【参考文献】