Unsupervised pavement anomaly detection via conditional diffusion model and domain adaptive feature selection

Pavement defects pose serious threats to road safety and infrastructure longevity. Following a full-supervised manner, many existing detection methods rely heavily on extensive labeled data. In this paper, motivated by the inherent diversity and imbalance of real-world pavement images, we propose a reconstruction-based unsupervised pavement anomaly detection framework. It leverages a conditional guided blurring diffusion model to reconstruct abnormal images as defect-free, combined with domain-adaptive feature refinement and a defect-aware feature selection module for robust anomaly scoring. By integrating simplex noise within the conditional guiding framework, our approach effectively preserves normal pavement textures while removing defects, enabling precise localization without relying on pixel-level annotations. Extensive comparison and ablation experiments on the Pavementscape dataset demonstrate that our method outperforms other unsupervised anomaly detection techniques and remains competitive with fully supervised segmentation approaches. These results underscore the potential of our unsupervised, diffusion-driven pipeline to address the costly annotation bottleneck in large-scale pavement inspection, offering a scalable and highly accurate solution for real-world road maintenance.