基于迁移学习与结构约束的无砟轨道裂缝分割算法
A Transfer Learning and Structural Constrained Algorithm for Ballastless Track Crack Segmentation
何梦欣 1黄华1
作者信息
- 1. 北京交通大学计算机科学与技术学院，北京 100044
- 折叠
摘要
针对无砟轨道裂缝检测中存在的数据规模有限、背景噪声干扰严重以及微细裂缝边缘提取困难等挑战，本文提出了一种融合迁移学习、语义引导边缘增强（SGEE）模块与结构相似性（SSIM）损失的深度学习框架。该框架通过在公开大规模数据集上进行预训练，并结合私有数据集微调，实现了模型从通用视觉特征向特定轨道场景的表征迁移，有效抑制了有限数据规模下的过拟合风险。实验结果表明，在无砟轨道私有数据集 GuiDaoBan 上，本文策略使F1分数提升了 9.31 个百分点，召回率从 30.05% 提升至 38.70%。该方法在复杂背景下展现出较强的边缘提取能力与稳健性，为数据受限场景下的轨道病害检测提供了良好的解决方案。
Abstract
To address the challenges of limited data scale, severe background noise interference, and the difficulty of extracting micro-crack edges in ballastless track crack detection, this paper proposes a deep learning framework integrating transfer learning, a Semantic-Guided Edge Enhancement (SGEE) module, and Structural Similarity (SSIM) loss. By pre-training on large-scale public datasets and fine-tuning on a private dataset, the framework facilitates the representation transfer of the model from general visual features to specific track scenarios, effectively suppressing the overfitting risk associated with limited data scale. Experimental results on the GuiDaoBan private ballastless track dataset demonstrate that the proposed strategy improves the F1-Score by 9.31 percentage points and increases the recall rate from 30.05% to 38.70%. The method exhibits superior edge extraction capability and robustness in complex backgrounds, providing an effective solution for track distress detection in data-constrained scenarios.关键词
裂缝检测/迁移学习/结构约束Key words
Crack detection/Transfer learning/Structural constraint引用本文复制引用
何梦欣,黄华.基于迁移学习与结构约束的无砟轨道裂缝分割算法[EB/OL].(2026-02-06)[2026-02-08].http://www.paper.edu.cn/releasepaper/content/202602-44.学科分类
铁路运输工程
评论