A robust defect detection method for catenary components based on structure-guided denoising diffusion model

Abstract

Ensuring the safety of railway operation requires accurate inspection of catenary components. Due to the lack of sufficient defective samples in practice, image reconstruction error-based methods have been developed for catenary inspection. However, these methods will encounter the problem that the reconstructed components may experience structural change or be blurry, which leads to lots of false detections. To tackle these issues, this article proposes a robust and reliable defect detection method called the catenary component structure-guided denoising diffusion model (CCSDM). The CCSDM consists of three key steps: 1) the recent denoising diffusion model is introduced to translate any component image to an image of normal component progressively, which ensures the reconstructed image with high quality; 2) to address structural change, we conditioned the component denoising process by constraining the reconstructed component to share the same low-frequency features as the input component, which ensures that they have similar structures; and 3) a twin-CNNs is proposed to localize the defective region by comparing the feature differences between the test sample and its reconstructed sample. The proposed CCSDM is a single-class detection approach that is applicable to different classes of components. In the experiment, CCSDM was evaluated on the dataset that includes six types of catenary components collected from China's high-speed rail line. The results show that CCSDM was able to reconstruct any test component image to its corresponding normal pattern image clearly and stably, which greatly reduces false detections and improves the accuracy of defect localization. Compared with the existing competitive methods, it achieves state-of-the-art performance with area under curve (AUC) exceeding 0.97 for different catenary components.