Crack detection from surface images using conditional diffusion models

Automated surface crack detection has become an essential task across various fields such as structural health monitoring, industrial inspection, material degradation analysis, and visual quality control systems. The ability to accurately identify cracks at an early stage not only mitigates safety risks but also significantly reduces long-term maintenance costs through timely interventions. However, detecting cracks often characterized by low contrast, irregular shapes, narrow widths, and sparse occurrences remains a persistent challenge in computer vision. To address this issue, we propose model, a novel segmentation framework that overcomes the inherent limitations of both supervised and unsupervised methods by adopting a conditional diffusion-based generative denoising approach. The model formulates crack mask prediction as a stepwise refinement process, where each prediction stage corresponds to a diffusion time step, enabling finer structural details to be progressively restored. Extensive experiments were conducted on two widely used benchmark datasets, Crack500, GAPs384, CrackForest, DeepCrack and CrackTree200, both containing diverse surface textures and crack variations. The results demonstrate that proposed model achieves significant performance gains over state-of-the-art methods in terms of Intersection over Union (IoU) and mean Dice Similarity (mDS). The model is particularly effective at capturing subtle, low-visibility cracks, highlighting its robustness in real-world, noisy conditions. Beyond improving segmentation accuracy, proposed model offers a scalable and generalizable solution applicable to a wide range of domains—from pavement surfaces and concrete walls to production lines and laboratory-grade material testing. As such, this study presents a reliable and high-capacity framework for automated crack detection, contributing to the advancement of modern inspection technologies and intelligent infrastructure monitoring systems.