Pavement crack segmentation with deep orthogonal-enhanced generative model

As high-resolution road-surface imagery provides a dynamic digital twin of pavement conditions, accurate crack segmentation serves as a key step in constructing intelligent maintenance systems. However, most existing segmentation models assume uniform sampling conditions and rely on fixed parameters, limiting their generalization under diverse real-world environments. To address this challenge, we propose a Deep Orthogonal-Enhanced Generative Model (DORGM) for robust pavement crack segmentation. The proposed framework introduces two key innovations: (1) an orthogonal constraint module that enforces feature disentanglement in the latent space, separating condition-specific noise from intrinsic crack patterns to reduce interference; and (2) a soft-label routing mechanism that adaptively assigns samples to specialized pathways, capturing subtle distributional shifts through Bayesian clustering within a variational framework. These modules can be seamlessly integrated into existing segmentation architectures, improving adaptability without additional retraining. Experiments conducted on benchmark datasets including DeepCrack, CRACK500, CFD, and NHA12D show consistent improvements — approximately 5% increase in mIoU and mDice — over baselines such as PSPNet, DNLNet, PointRend, SegFormer, and VPD. By enforcing orthogonality in the latent space, DORGM effectively disentangles environmental variations (e.g., lighting or weather-induced noise) from core crack features, yielding stable and interpretable segmentation across heterogeneous data sources. This disentanglement further mitigates domain shifts that hinder digital twin applications, enabling more reliable integration of crack data with other urban layers such as traffic flow and structural health metrics.