面向精确定位的列车车号文本定位与识别  

Precise Localization and Recognition of Train Characters

作　　者：艾鑫 邹琪[1] 罗常津 Ai Xin;Zou Qi;Luo Changjin(Department of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044;Institute of Electronic Computing Technology,China Academy of Railway Sciences,Beijing 100081)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]中国铁道科学研究院电子计算技术研究所,北京100081

出　　处：《计算机辅助设计与图形学学报》2020年第11期1863-1870,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61473031);中央高校基本科研业务费专项(2019JBM019).

摘　　要：针对现有文本定位方法应用于列车车号易出现定位边界不紧凑、面积占比小的车号漏检率高的问题,提出一种面向小尺度目标精确定位的列车车号定位与识别方法.其基于CTPN改进,在文本定位阶段,首先采用VGG16网络提取特征并融合多尺度的特征图,以利于定位小车号区域;其次采用区域建议网络生成候选区域,对其进行分类回归,分类过程设计了困难样本挖掘策略,即保留只包含半个数字的正样本,回归过程设计了边界敏感的细粒度文本框回归策略,以确保水平边界紧凑;最后连接候选区域,输出定位结果.文本识别阶段采用基于注意机制的文本识别方法.通过在Caffe环境验证车号检测数据集,结果表明,车号定位方法相对经典的文本定位方法提高0.11,车号整体识别F1分数为0.81.To solve the problems of the non-compact boundary localization and the loss of small objects,a train number localization and recognition method especially for precise localization of small-scale targets is proposed.This method improves from CTPN in the text localization stage.We first use the VGG16 network to extract features and fuse multi-scale feature maps,which is beneficial to small text areas and then use the region proposal network to generate candidate areas,for classification and regression.A hard sample mining strategy,that is,to retain positive samples containing only part of the train number is designed in the classification process.A boundary-sensitive fine-grained text box regression strategy in the regression process is designed to ensure that the horizontal boundary is compact.Finally the candidate regions are connected to output the localization results.In the text recognition stage,a text recognition method based on attention mechanism is used.By testifying the train number detection data set in the Caffe environment,the results show that the train number localization method is 0.11 higher than the classic text localization method.The overall train number recognition F1 value is 0.81.

关 键 词：车号定位 车号识别 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]