计算机与现代化 ›› 2024, Vol. 0 ›› Issue (06): 89-94.doi: 10.3969/j.issn.1006-2475.2024.06.015

• 图像处理 • 上一篇    下一篇

小样本下基于改进度量学习的轨面状态识别#br#

  

  1. (1.湖南工业大学轨道交通学院,湖南 株洲 412007; 2.湖南铁道职业技术学院智能控制学院,湖南 株洲 412012)
  • 出版日期:2024-06-30 发布日期:2024-07-17
  • 作者简介: 作者简介:于惠钧(1975—),男,河南驻马店人,教授,硕士生导师,博士,研究方向:图像处理,系统保护与自动化技术,测试与故障诊断,自动控制,E-mail: arejunyu@foxmail.com; 彭慈兵(2000—),男,湖南宁乡人,硕士研究生,研究方向:小样本学习,图像处理,E-mail: bigeral@qq.com; 通信作者:刘建华(1981—),男,河南商水人,副教授,硕士生导师,博士,研究方向:轨道交通智能运维,模式识别,电传动与控制,E-mail: jhliu0615@163.com; 张锦圣(1997—),河南驻马店人,硕士研究生,研究方向:深度学习,图像处理,E-mail: 1114788239@qq.com; 刘丽丽(1985—),女,河南驻马店人,副教授,研究方向:复杂系统建模与控制,图像处理,E-mail: 18018999851@163.com。
  • 基金资助:
    国家自然科学基金资助项目(52272347); 湖南省自然科学基金资助项目(2021JJ30217,2022550095); 湖南省教育厅科学研究重点项目(20A162)
      

Rail Surface State Identification Based on Improved Metric Learning under Small Samples



  1. (1. School of Rail Transportation, Hunan University of Technology, Zhuzhou 412007, China;
    2. School of Intelligent Control, Hunan Railway Professional Technology College, Zhuzhou 412012, China)
  • Online:2024-06-30 Published:2024-07-17

摘要:
摘要:为解决小样本下轨面状态识别过程中存在的关键特征信息提取不充分、区分度信息易丢失的问题,提出一种基于改进度量学习的轨面状态识别方法。该方法在特征提取网络部分引入金字塔拆分注意力机制,实现特征图空间信息多尺度提取、跨维度通道注意力与空间注意力特征交互,以解决轨面状态样本少导致的关键特征信息提取不充分的问题。利用深度局部拼接符对查询集与各类支撑集特征图进行局部特征两两拼接,代替传统度量学习的全局特征拼接,筛选背景等干扰信息,较大程度地保留有显著区分度的特征信息。在自建小样本轨面状态数据集上进行性能验证,并与常规小样本学习方法进行对比实验,实验结果表明,本文方法能够有效识别轨面状态,识别准确率、精度、召回率、F1值分别达到97.96%、98.61%、98.07%、98.34%,相比于性能较好的小样本学习方法DN4网络,各项指标分别提升了5.75个百分点、5.83个百分点、5.95个百分点、5.89个百分点。

 


关键词: 关键词:轨面状态识别, 小样本, 度量学习, 金字塔拆分注意力, 深度局部拼接

Abstract: Abstract: In order to solve the problems of insufficient extraction of key feature information and easy loss of discrimination information in the process of rail surface state identification under small sample conditions, a rail surface state identification method based on improved metric learning is proposed. This method incorporates a pyramid split attention mechanism in the feature extraction network to achieve multi-scale extraction of feature map spatial information, cross-dimensional channel attention and spatial attention feature interaction, so as to solve the problem of insufficient extraction of key feature information caused by the small number of track state samples. Additionally, a deep local splicing operator is employed to splice the local features of the query set and various support set feature maps in pairs, replacing the global feature splicing used in traditional metric learning. This helps fitter out filtering interference information such as background noise, and retains significant distinguishing feature information to a greater extent. Experimental results show that the proposed method can effectively identify the rail surface status, and the recognition accuracy, precision, recall rate, and F1 value reach 97.96%, 98.61%, 98.07%, and 98.34%, respectively. Compared with the small sample learning method the DN4 network with better performance, these indicators increased by 5.75, 5.83, 5.95, and 5.89 percentage points, respectively.

Key words: Key words: rail surface state recognition, small sample, metric learning, pyramid split attention, deep local splicing

中图分类号: