中国铁道科学2017，Vol.38Issue(3)：89-96,8.DOI:10.3969/j.issn.1001-4632.2017.03.13
基于深度学习的铁道塞钉自动检测算法
Automatic Detection Algorithm for Rail Plug Based on Deep Learning
摘要
Abstract
We propose automatic detection algorithms for rail plug using deep learning methods based on some classical algorithms in service.In consideration of some specialties of these plug images that are captured by cameras installed on the high-speed railway inspection vehicle,we propose the spectrum residual region proposal (SRP) method to obtain proposal plug patches in the stage of regional selection and design the plug convolution neural network (pCNN) in the stage of feature extraction.Inspired by salient detection,after comparing the spectrum difference between plug images and average images without plugs,we use FFT to obtain residual spectrums of those images and IFFT to get some proposal regions.After the procession of SRP mentioned above,we make use of pCNN to extract features layer by layer and obtain some feature images that imply intrinsic specialties of plug patches.Specially,in the structure of pCNN,there are 4 convolution layers,3 pooling layers,3 nonlinear transformation layers,3 normalized layers,2 full connected layers and 1 drop layer,and its outputs are image features that are used as inputs of SVM in the following procession.Finally,we use SVM to determine whether these patches are plug regional images and thus the automatic location of plug can be realized.Comparing with other methods,our method achieves state of the art results in lots of practical tests.关键词
塞钉/轨道电路/目标检测/图像识别/深度学习/卷积神经网络/高速铁路Key words
Plug/Track circuit/Object detection/Image recognition/Deep learning/Convolution neural network/High speed railway分类
交通工程引用本文复制引用
杜馨瑜,戴鹏,李颖,程雨,王胜春,韩强,王昊..基于深度学习的铁道塞钉自动检测算法[J].中国铁道科学,2017,38(3):89-96,8.基金项目
国家“九七三”计划项目(2013CB329400) （2013CB329400）
中国铁路总公司科技研究开发计划重大项目(2015T003-A) （2015T003-A）
中国铁道科学研究院行业服务技术创新项目(2014YJ052) （2014YJ052）
