Abstract: Railway track foreign object detection plays a crucial role in ensuring the safe operation of railways. However, this field currently faces two major challenges: data scarcity and high annotation costs. Some anomalies on the tracks are relatively rare, making it difficult for existing public datasets to cover a wide range of abnormal scenarios. Furthermore, manual data annotation is not only time-consuming and labor-intensive but also struggles to meet the demands of large-scale applications. To address these challenges, this paper proposes a novel framework for railway track foreign object image generation and detection, combining foreign object image generation with semi-supervised learning strategies to enhance detection performance. Specifically, the paper introduces a multi-region guided foreign object generation method based on a diffusion model, which can simultaneously generate realistic railway foreign object images across multiple regions while maintaining overall stylistic consistency. This approach effectively mitigates the problem of insufficient real data. Additionally, the paper develops a semi-supervised detection framework based on cross-view query consistency, which addresses the issue of noisy pseudo-labels in the teacher-student framework by learning more robust semantic features across different augmented views. Extensive experimental results demonstrate that the proposed method significantly improves both the accuracy and robustness of track foreign object detection, providing an efficient and reliable solution for safe railway operations.

Key words: track obstacle detection, image generation, semi-supervised learning

摘要： 铁路轨道异物检测在保障铁路正常运营方面具有重要意义。然而，目前该领域主要面临两大挑战：数据稀缺和标注成本高。由于轨道上的某些异常较为罕见，现有公开数据集难以覆盖多样化的异常情况；而人工标注数据不仅耗时费力，且难以满足大规模应用需求。为了应对这些挑战，提出了一种新颖的铁路轨道异物图像生成与检测框架，结合异物图像生成和半监督学习策略以提升检测性能。具体而言，针对数据稀缺问题，提出了一种基于扩散模型的多区域引导异物生成方法，能够在多个区域同时生成逼真的铁路异物图像，并保持整体风格的一致性，从而有效缓解真实数据不足的问题。此外，为了解决标注成本高的问题，还构建了一种基于跨视图查询一致性的半监督检测框架，通过在不同增强视图之间学习更加稳健的语义特征，有效地解决了教师-学生框架中噪声伪标签问题。大量实验结果表明，提出的方法在轨道异物检测任务中显著提升了精度与鲁棒性，为铁路安全运营提供了一种高效可靠的解决方案。

关键词: 轨道异物检测, 图像生成, 半监督学习