NNPSQE: framework for enhancing quality-based phase unwrapping via neural-network phase separation

Quality-based phase-unwrapping methods depend critically on the reliability of quality maps. However, quality metrics computed directly from the original wrapped phase often respond similarly to valid rapid phase variations and genuine unreliable factors. To reduce this ambiguity, we propose NNPSQE (neural-network phase-separation quality enhancement), a framework for enhancing quality-based phase unwrapping via neural-network phase separation. NNPSQE does not use the neural network either as a conventional phase-fitting module or as a direct phase-unwrapping network. Instead, the network estimates a constrained continuous trend phase as an intermediate variable for quality-map enhancement. During network optimization, the trend phase is obtained by explicitly minimizing the degradation response of the same quality metric used by the downstream unwrapping solver, where the response is computed from the optimization-step residual wrapped phase generated at each optimization step. Within the constrained trend model, degradation caused by valid rapid phase variations is encouraged to be explained by the trend phase, whereas genuine unreliable components are discouraged from being absorbed. Specifically, a deep decoder prior, together with Laplacian smoothness and gradient-barrier constraints, restricts the trend phase to continuous and gradient-bounded components. After trend removal, the quality metric is recomputed in the final residual wrapped-phase domain. This residual-domain quality re-evaluation substantially alleviates the original-domain ambiguity between valid rapid phase variations and genuine unreliable factors, thereby improving the reliability of the quality map. The downstream unwrapping process is left unchanged, so the framework can serve as a flexible front-end for phase-unwrapping methods based on quality maps. Experiments on simulated, cell, film, 3D-printed surface, fringe-projection structured-light 3D, and InSAR phase data demonstrate that the proposed framework improves the accuracy and reliability of quality-based phase unwrapping and shows strong cross-data generalization without pretraining or labeled data.