RGB-D Rail Surface Defect Inspection Driven by Conditional Diffusion Architecture and Frequency Knowledge

RGB-D rail surface defect inspection (RSDI) is a critical measure for ensuring transportation safety. It improves inspection accuracy by using depth maps, but the issue of poor-quality depth maps in rail defect datasets is often overlooked. Additionally, the similarity between the foreground and the background has always been a persistent challenge in RGB-D RSDI. In this article, we propose a novel network, termed DiffRSDI to overcome these challenges in RGB-D RSDI, which treats RGB-D RSDI as a mask generation task using a conditional diffusion architecture. Specifically, DiffRSDI consists of two key components: a cross-modal cyclic enhancement (CCE) module and a frequency-aware alternate fusion (FAF) module. The CCE module is based on the cyclic enhancement approach and aims to eliminate interference from poor-quality depth maps and aggregate cross-modal features. The FAF module overcomes the similarity between the foreground and the background in the frequency domain and achieves cross-scale fusion in the fused features. Comprehensive experiments conducted on the NEU RSDDS-AUG dataset confirm that our method achieves superior performance in comparison to 14 related methods. The DiffRSDI code and results are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/zeroyi37/DiffRSDI</uri>.