Local structure parsing and reconstruction of road and bridge cracks based on a conditional latent diffusion model

Abstract

Accurate crack segmentation on road and bridge surfaces underpins automated structural-health assessment, yet field images are routinely degraded by occluding stains, repaired patches, cast shadows, blur and noise. Under such degradation, discriminative networks tend to fracture a slender, low-contrast crack into disconnected fragments, corrupting downstream measurements of crack length, width and connectivity. We recast crack extraction as conditional generative reconstruction and propose CrackDiff, a conditional latent diffusion model guided by an explicit Local Structure Parsing (LSP) descriptor that encodes multiscale ridge orientation, a soft skeleton and a centreline-distance saliency. A KL-regularised autoencoder compresses the complete crack map into a compact latent space, where a conditional U-Net denoiser reconstructs the full crack topology from the degraded image and its LSP descriptor; training augments the noise-prediction loss with a topology-aware soft-clDice term that rewards connectivity. On the DeepCrack benchmark under a controlled degradation protocol, CrackDiff reconstructs continuous crack centrelines and attains a high clDice score, and across four degradation strengths its region and topology metrics decline only gradually. These results show that pairing an explicit local-structure prior with a topology-aware generative objective is an effective route to robust crack parsing under real-world degradation.