Memory‐Guided Multi‐Constraint Deep Unrolling for Robust Phase Unwrapping

Phase unwrapping (PU) aims to recover continuous phase maps from wrapped measurements and is fundamental to graphics, depth sensing, holography, and remote sensing. However, accurate PU remains challenging due to noise, discontinuities, and severe wrapping ambiguities. In this paper, we propose MM‐DPU , a memory‐guided multi‐constraint deep unfolding framework for robust phase unwrapping. The iterative optimization process is unfolded into a stage‐wise reconstruction network, where each stage alternates between wrapped‐gradient data consistency and learned phase‐prior refinement. To improve error correction across stages, we introduce a dual residual memory mechanism that propagates the discrepancy between the data‐consistent estimate and the phase‐prior output. We further enhance the phase‐prior network with a dual‐branch bottleneck, combining lightweight global‐context modeling and spatial‐gated modulation to emphasize informative phase structures and suppress ambiguous wrapped patterns. The network is trained with a hybrid objective that integrates unsupervised wrapped‐gradient fidelity and supervised phase‐gradient constraints, jointly improving physical consistency and local structural accuracy. Experiments on PU for time‐of‐flight imaging, compressive holographic video, and Interferometric Synthetic Aperture Radar demonstrate the robustness and generalization ability of our method across diverse phase unwrapping scenarios.