Structured light 3D imaging via fusion of spatial-frequency information and domain-adaptive adversarial learning

Structured-light 3D imaging plays a pivotal role in high-precision, non-contact measurement; however, current deep learning methods exhibit limited generalization across varying system configurations and environmental conditions. To address these challenges, this paper proposes a structured-light 3D imaging framework integrating domain-adaptive adversarial learning and a dual-domain attention mechanism. Firstly, we design an innovative dataset generation algorithm capable of simulating diverse imaging conditions-including variations in material properties, illumination environments, fringe frequencies, and system configurations-by combining real-world fringe measurements and physically accurate illumination models. Secondly, we introduce a novel attention network, termed FMambaBlock, which effectively fuses spatial-domain attention via a two-dimensional state-space model (2D-SSM) and frequency-domain amplitude-phase attention, significantly enhancing the representation of global and local features critical for fringe pattern analysis. Additionally, we incorporate domain adversarial training to explicitly learn domain-invariant features, thereby substantially improving the model's robustness and generalization to unseen domains. Extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in cross-domain scenarios involving unknown fringe frequencies, illumination variations, and system parameter perturbations, significantly outperforming existing approaches and proving its practical applicability.