CrackDiffusion: a two-stage semantic segmentation framework for pavement crack combining unsupervised and supervised processes

Abstract

Achieving precise and reliable automated pavement crack detection using deep learning techniques is vital for intelligent pavement maintenance. This study proposes CrackDiffusion, an enhanced-supervised detection framework for pavement crack, combining two supervised and unsupervised stages. In Stage 1, a multi-blur-based cold diffusion anomaly detection model is proposed, which transforms crack-containing images into crack-free images, while simultaneously extracting pixel-level crack features using the Structural Similarity Index measure (SSIM). In Stage 2, an improved supervised U-Net segmentation model enhances accuracy and robustness by building upon the unsupervised results from Stage 1, ultimately producing highly accurate pixel-level segmentation results for cracks. On four public datasets, both the proposed multi-blur-based cold diffusion model and the comprehensive CrackDiffusion framework attained the highest Intersection over Union (IoU) scores, surpassing the IoU scores of the current state-of-the-practice unsupervised and supervised segmentation models.