CrackDiffusion: crack inpainting with denoising diffusion models and crack segmentation perceptual score

Abstract

Abstract Cracks commonly occur in engineering structures. Imaging method is one of the most effective detection method for crack. However, crack information captured by the imaging sensors is often interfered by noise and the other environmental factors. In this paper, we propose a crack inpainting method that can automatically repair the missing crack information. The inpainting method consists of a denoising diffusion model and a segmentation guidance model. Taking advantages of denoising diffusion model’s stability and segmentation guidance model’s accuracy, we can achieve coherent inpainting patches as well as accurate crack traces. Furthermore, we propose a fine crack metric—crack segmentation perceptual score to guide high quality crack generation. Experimental results show that our method achieves both high quality and precise crack inpainting results, which is very beneficial to the crack detection and evaluation in structural health monitoring.