测绘学报 ›› 2026, Vol. 55 ›› Issue (10): 1878-1891.doi: 10.11947/j.AGCS.2026.20260170

• 地图学与地理信息 • 上一篇    下一篇

多阶段特征协同增强的铁路轨道构件缺陷高效检测方法

谢亚坤1,2(), 柏雯静1, 曹奋扬1, 刘杨苹1, 朱军1,2(), 朱庆1   

  1. 1.西南交通大学地球科学与工程学院,四川 成都 611756
    2.西南交通大学桥梁智能与绿色建造全国重点实验室,四川 成都 611756
  • 收稿日期:2026-05-08 修回日期:2026-09-20 出版日期:2026-11-16 发布日期:2026-11-16
  • 通讯作者: 朱军 E-mail:yakunxie@swjtu.edu.cn;zhujun@swjtu.edu.cn
  • 作者简介:谢亚坤(1991—),男,博士,副教授,研究方向为铁路数字孪生建模。E-mail:yakunxie@swjtu.edu.cn
  • 基金资助:
    国家重点研发计划(2023YFB2603700);国家自然科学基金(42301473; 42671553);四川省科技计划资助(2026NSFSCZY0054; 2025ZNSFSC0322);成都市科技计划项目(2025-YF05-00009-SN);山东铁投集团科技计划(TTKJ2025-07)

A multi-stage feature collaborative enhancement method for efficient defect detection in railway track components

Yakun Xie1,2(), Wenjing Bai1, Fenyang Cao1, Yangping Liu1, Jun Zhu1,2(), Qing Zhu1   

  1. 1.Faculty of Geosciences and Engineering, Southwest Jiaotong University, Chengdu 611756, China
    2.State Key Laboratory of Bridge Intelligent and Green Construction, Southwest Jiaotong University, Chengdu 611756, China
  • Received:2026-05-08 Revised:2026-09-20 Online:2026-11-16 Published:2026-11-16
  • Contact: Jun Zhu E-mail:yakunxie@swjtu.edu.cn;zhujun@swjtu.edu.cn
  • About author:Xie Yakun (1991—), male, PhD, associate professor, majors in railway digital twin modeling. E-mail: yakunxie@swjtu.edu.cn
  • Supported by:
    The National Key Research and Development Program of China(2023YFB2603700);The National Natural Science Foundation of China(42301473; 42671553);Sichuan Science and Technology Program(2026NSFSCZY0054; 2025ZNSFSC0322);Chengdu Science and Technology Program(2025-YF05-00009-SN);Science and Technology Program of Shandong Railway Investment Holding Group Co., Ltd.(TTKJ2025-07)

摘要:

铁路轨道构件缺陷检测对线路服役安全与智能养护具有重要意义。针对轨道巡检影像中小尺度构件细节易衰减、局部缺陷特征弱、复杂背景干扰强及相似构件状态易混淆等问题,本文提出一种多阶段特征协同增强的铁路轨道构件缺陷高效检测方法。该方法围绕鱼尾板孔位边缘、构件连接区、紧固件轮廓及缺失邻域等结构判别线索,构建由细节保持、特征协同和判别聚焦组成的递进式增强机制。在特征提取阶段,采用内容自适应下采样模块动态调整特征压缩响应,减少小尺度结构信息损失;在特征融合阶段,利用多源特征均衡模块协调不同层级和传播路径的特征表达,增强局部细节与高层语义的协同关系;在检测输出前,通过大核自适应注意力模块强化关键区域响应,提高复杂背景下相似构件状态的区分能力。同时,构建包含1528张图像和2367个目标实例的铁路轨道构件缺陷检测数据集。试验结果表明,本文方法的Precision、Recall、mAP@0.5和mAP@[0.5∶0.95]分别达到91.4%、87.3%、90.1%和68.9%,参数量为4.6×106,计算量为7.0×109 FLOPs,推理速度为101.0帧/s。复杂扰动、跨数据集泛化、稳定性分析和消融试验进一步验证了本文方法的有效性。

关键词: 铁路轨道, 缺陷检测, 多阶段特征, 协同增强

Abstract:

Defect detection in railway track components is of great significance for railway operational safety and intelligent maintenance. To address the problems of detail degradation in small-scale components, weak local defect features, strong background interference, and confusion among visually similar component states in track inspection images, this paper proposes an efficient defect detection method for railway track components based on multi-stage feature collaborative enhancement. The proposed method focuses on structural discriminative cues such as fishplate hole edges, component connection regions, fastener contours, and missing-neighborhood structures, and constructs a progressive enhancement mechanism consisting of detail preservation, feature collaboration, and discriminative focusing. In the feature extraction stage, a content-adaptive downsampling module is adopted to dynamically adjust the feature compression response and reduce the loss of small-scale structural information. In the feature fusion stage, a multi-source feature equalization module is used to coordinate feature representations from different levels and propagation paths, thereby enhancing the collaboration between local details and high-level semantics. Before detection output, a large adaptive kernel attention module is employed to strengthen key-region responses and improve the discrimination of similar component states under complex backgrounds. Meanwhile, a railway track component defect detection dataset containing 1,528 images and 2,367 object instances is constructed. Experimental results show that the proposed method achieves Precision, Recall, mAP@0.5, and mAP@[0.5: 0.95] values of 91.4%, 87.3%, 90.1%, and 68.9%, respectively, with 4.6×106 parameters, 7.0×109 FLOPs, and an inference speed of 101.0 FPS. Experiments on complex perturbations, cross-dataset generalization, stability analysis, and ablation studies further verify the effectiveness of the proposed method.

Key words: railway track, defect detection, multi-stage feature, collaborative enhancement

中图分类号: