SG-AMDM: a semantic-guided adaptive mask diffusion model for catenary component defect detection in electrified railways

Abstract

In electrified railways, health condition monitoring and measurement of Catenary Support Components (CSCs) are essential for safe operation. However, existing methods rely on category-specific templates or sufficient labelled defect samples. To address these limitations, this paper proposes a novel unsupervised detection framework: the Semantic-Guided Adaptive Mask Diffusion Model (SG-AMDM). Firstly, we propose a novel Adaptive Mask Generator (AMG) to generate adaptive masks, thereby guiding the inpainting network to reconstruct potential defects. Secondly, we propose a Semantic-Guided Diffusion Model to reconstruct the defect region using global context-semantic information consistent with the normal structure, thereby avoiding category interference during multi-class reconstruction. Furthermore, a Spatial-aware Feature Fusion block is integrated to handle CSCs’ scale variations and effectively maximize reconstruction accuracy. Experiments on a real multi-class CSC dataset collected by the 4C inspection vehicle demonstrate that SG-AMDM achieves detection performance of 99.7/99.7/99.5% in image-level metrics (I-AUROC/AP/F1) and 98.3/38.4/57.8% in pixel-level metrics (P-AUROC/AP/F1).