基于深度学习的铁道塞钉自动检测算法

发布时间：2018-05-14 09:38

  本文选题：塞钉 + 轨道电路　； 参考：《中国铁道科学》2017年03期

【摘要】：根据高铁巡检车所采集轨腰图像中铁道塞钉图像的特点,在既有计算机视觉的目标检测算法的基础上,提出基于深度学习的铁道塞钉自动检测算法。在目标检测的区域选择阶段,借鉴显著性检测的思路,提出余谱区域候选(Spectrum Residual Region Proposal,SRP)算法,即利用含塞钉的轨腰图像与不含塞钉的轨腰平均图像之间的频谱差异,通过快速傅里叶变换,得到两图像间的幅度谱差的绝对值(余谱),再通过快速傅里叶反变换及后处理,得到候选目标区域;然后在目标检测的特征提取阶段,设计塞钉卷积神经网络(plug Convolution Neural Network,pCNN),该网络通过4个卷积层、3个池化层、3个非线性变换层、3个规范化层、2个全连接层和1个泄露层,自动从候选目标区域逐层提取最能表现塞钉特征的特征图像;最后基于特征图像采用支持向量机(SVM)的分类器判断候选目标区域是否含有塞钉,从而实现塞钉的自动定位。大量实际测试以及与其他算法比较的结果表明,该算法的检测效果最优。
[Abstract]:According to the characteristics of railway stud images collected by high-speed railway inspection vehicle, an automatic detection algorithm based on depth learning is proposed on the basis of the existing target detection algorithms of computer vision. In the region selection stage of target detection, using the idea of significant detection for reference, this paper proposes a candidate Spectrum Residual Region Proposal Residual Region algorithm for cospectral region, that is, using the spectral difference between the rail waist image with studs and the average rail waist image without studs. The absolute value of amplitude spectral difference between two images (cospectrum) is obtained by fast Fourier transform (FFT), and then the candidate target region is obtained by FFT and post-processing, and then in the feature extraction stage of target detection, Plug Convolution Neural network pCNNs are designed. The network consists of four convolution layers, three pool layers, three nonlinear transformation layers, three normalized layers, two fully connected layers and one leak layer. Finally, the feature image is extracted from the candidate target area layer by layer. Finally, the support vector machine (SVM) classifier is used to determine whether the candidate target region contains studs or not, so as to realize the automatic location of studs. A large number of practical tests and comparison with other algorithms show that the algorithm has the best detection effect.
【作者单位】： 中国铁道科学研究院基础设施检测研究所;
【基金】：国家“九七三”计划项目(2013CB329400) 中国铁路总公司科技研究开发计划重大项目(2015T003-A) 中国铁道科学研究院行业服务技术创新项目(2014YJ052)
【分类号】：TP18;U216.3
，

本文编号：1887332