HMSPU-Net: hierarchical multi-scale state space model with physics-guided refinement for high-accuracy phase unwrapping in complex scenes

Le résumé fourni par la source

Dense fringes, noise, and surface discontinuities persistently cause large-scale fringe-order errors in deep-learning phase unwrapping, as existing methods lack long-range context and physical constraints. We address this with HMSPU-Net, a hierarchical multi-scale state-space model with physics-guided residual refinement. Efficient 2D Scanning captures global context at linear complexity, while coarse-to-fine state fusion preserves sharp phase-jump boundaries. A wrapped-residual head and an adaptive gated loss embed physical priors. HMSPU-Net achieves an NRMSE of 1.53% and an AU of 97.34% under severe noise, and an AU of 98.45% on the self-constructed real-world dataset, outperforming existing methods and demonstrating robust phase unwrapping in complex scenes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.