MSPANet: a multi-scale phase-aware network for spatial phase unwrapping via fringe order classification

This paper presents a multi-scale phase-aware network (MSPANet) for spatial phase unwrapping in structured-light three-dimensional (3D) reconstruction. Fringe-order estimation is formulated as a semantic segmentation problem to avoid error accumulation and integer constraints inherent in regression-based approaches. Guided by the physical characteristics of wrapped-phase propagation, MSPANet integrates parallel multi-scale feature extraction with dilated convolution, learnable edge enhancement, and a phase-aware gating module to handle periodic 2π discontinuities. A hybrid loss function incorporating continuity, edge alignment, and regional consistency further enforces physically plausible predictions. Extensive experiments on simulated and real-world datasets demonstrate that MSPANet achieves robust and accurate phase unwrapping under noise and discontinuities, supporting reliable 3D shape measurement in complex scenarios.