Noise-robust phase unwrapping for smooth phase fields with physics-consistent constraints

Phase unwrapping (PU) aims to recover a continuous absolute phase (AP) from a wrapped phase (WP) and plays a fundamental role in quantitative optical measurement and computational imaging. In practical scenarios, measurement noise often disrupts phase continuity and significantly degrades the robustness of conventional PU methods. However, many existing approaches either rely on heuristic regularization or directly learn the WP–AP mapping without explicitly incorporating physical constraints, leading to limited robustness and interpretability. In this work, we focus on noise-resistant phase unwrapping under smooth or moderately varying phase conditions, and propose PUN-LP, a physics-consistent PU network with an enlarged receptive field. PUN-LP integrates a visual state-space model to efficiently capture long-range spatial dependencies, together with a gated bottleneck convolution to adaptively model local phase variations. To ensure physical consistency, we further introduce a hybrid loss that combines multi-scale data-consistency supervision with explicit constraints derived from the forward phase wrapping model, thereby enforcing the learned mapping to conform to the underlying physical law. Extensive experiments on synthetic and real-world datasets demonstrate that PUN-LP achieves improved robustness and generalization in noisy environments compared with existing PU methods. Nevertheless, the current framework primarily addresses noise-induced phase distortions and does not explicitly model challenging cases involving sharp discontinuities or highly complex fringe structures, which will be investigated in future work.