Manifold regularization for Semantic Segmentation on Asphalt Pavement Images Via Adversarial Example

Timely and accurate detection of road surface distress and design features is essential for intelligent pavement surveys. Nevertheless, pavement images acquired in real-world surveys frequently exhibit high similarity between pavement features and non-target textures, along with acquisition-induced noise, both of which significantly hinder the detection accuracy and model robustness. Consequently, this paper introduces a deep learning model termed AD-Unet, which integrates adversarial example manifold regularization (ADMR) into the classical Unet architecture to tackle these challenges. In this framework, adversarial examples are generated based on the original pavement data while preserving the topological similarity in the manifold space. By constraining network training through ADMR, the proposed method enhances the model’s resilience and significantly improves segmentation accuracy under complex conditions. Experimental evaluations demonstrate that the proposed AD-Unet achieves a mean F1-Score of 78.65% and a mean intersection-over-union (IoU) of 70.94% on testing images, respectively. Furthermore, comparative studies conducted on both proprietary and public datasets illustrate that AD-Unet yields a noticeably higher detection accuracy than other semantic segmentation models for diverse pavement features.