基于PSENet区域分割的列车车厢号定位研究  

Research on Train Carriage Number Location Based on Region Segmentation Using PSENet

作　　者：范沐阳 喻春雨[1] 马鑫 童亦新 张俊[1] FAN Muyang;YU Chunyu;MA Xin;TONG Yixin;ZHANG Jun(College of Electronic and Optical Engineering&College of Flexible Electronics(Future Technology),Nanjing University of Posts and Telecommunications,Nanjing 210023,China)

机构地区：[1]南京邮电大学电子与光学工程学院、柔性电子(未来技术)学院,江苏南京210023

出　　处：《铁道学报》2024年第10期68-75,共8页Journal of the China Railway Society

基　　金：南京邮电大学产学研合作项目(KH1060319157、KH1060318002)。

摘　　要：针对自然场景中行驶列车的车厢号采用现有文本定位方法容易存在严重畸变,从而导致漏检率高的问题,提出一种基于区域分割的列车车厢号精确定位方法。该方法所采用的网络模型基于PSENet,首先采用引入通道域注意力的ResNet50网络提取特征,使模型更加关注卷积通道信息,从通道域的角度对权重进行再分配,提高车厢号的定位精度;然后采用特征金字塔和自底向上路径增强模块融合多尺度特征图,将浅层网络中的强定位特征传播到深层网络,以从复杂环境中准确定位车厢号区域,降低车厢号的漏检率;最后采用基于广度优先算法的渐进尺度扩展模块对融合的特征图进行目标区域的从小尺度到大尺度的拓展分割,使用结合集合相似度度量Dice系数的损失函数对分割结果进行分类回归,输出定位结果。实验结果表明:通过在真实列车车厢号图像数据集上进行训练验证,本文提出的网络模型在铁路货运列车车厢号定位精度达到97.47%,召回率为94.14%,综合F1分数为95.78%,预测单张图约需0.2 s。相比于基础PSENet对车厢号的定位精确率提升3.78%,召回率提升1.71%,总体上车厢号的F1分数提升2.73%。研究成果可为列车车厢号自动检测识别系统提供一种高精度的车厢号检测定位方法。Aiming at the problem of high missed detection rate caused by the serious distortion of the existing text location methods for the carriage number of running trains in natural scenes,an accurate location method was proposed for the carriage number of trains based on region segmentation.The network model used in this method is based on PSENet.First,the ResNet50 network with channel domain attention was used to extract features,to cause the model to focus more on convolution channel information,to redistribute the weights from the perspective of channel domain,so as to improve the positioning accuracy of carriage number.Then the feature pyramid and bottom-up path enhancement module were used to fuse multi-scale feature maps,where the strong positioning features in the shallow network were propagated to the deep network to accurately locate the carriage number area from the complex environment,which is conducive to reducing the missing rate of carriage numbers.Finally,the gradual scale expansion module based on the breadth first algorithm was used to expand and segment the fused feature map from small scale to large scale,while the loss function combined with the set similarity measure Dice coefficient was used to classify and regress the segmentation results and output the positioning results.The research results show that,through training and verification on the real train carriage number image dataset,the network model proposed in this paper achieves a positioning precision of 97.47%,a recall rate of 94.14%,a comprehensive F1 score of 95.78%,and a prediction time of about 0.2 seconds for a single image.Compared to the basic PSENet,the positioning accuracy of the carriage number has increased by 3.78%,with the increase of the recall rate by 1.71%.Overall,the F1 score of the car number has increased by 2.73%.The research results can provide a high-precision detection and location method for the automatic detection and identification system of train carriage numbers.

关 键 词：货运列车 车厢号定位 PSENet 通道注意力 多尺度特征融合 

分 类 号：TP394.41[自动化与计算机技术—计算机应用技术] U268.2[自动化与计算机技术—计算机科学与技术]