A two-stage phase unwrapping method by deep learning on noisy wrapped phases

Liang Wan, Jiayang Liu, Yi’nan Qian, Yong Kang, Shijing Wu, Deng Li,
A two-stage phase unwrapping method by deep learning on noisy wrapped phases,
Optics & Laser Technology,
Volume 192, Part D,
2025,
113787,
ISSN 0030-3992,
https://doi.org/10.1016/j.optlastec.2025.113787.
(https://www.sciencedirect.com/science/article/pii/S0030399225013787)
Abstract: Phase unwrapping is an essential part of modern measurement techniques including fringe projection profilometry, synthetic aperture radar, digital holographic interferometry, magnetic resonance imaging. Traditional phase unwrapping methods not only lead to path-dependent and over-smoothed results but also consume unaffordable time when dealing with large phase maps. Existing deep learning phase unwrapping methods lack sufficient accuracy when dealing with highly noise-contaminated wrapped phase maps, as they fail to fully exploit the inherent features of the wrapped phase maps. For this reason, a two-stage phase unwrapping method is proposed. The proposed network includes a denoise convolutional neural network (DnCNN), a preliminary phase unwrapping network (PPUN), and a fine phase unwrapping network (FPUN). Among them, the DnCNN of Stage-I can reduce the challenge of phase unwrapping while providing relatively clean gradient information for FPUN by minimizing the noise of wrapped phases. In Stage-II, convolutional block attention module (CBAM) is introduced into the PPUN based on semantic segmentation for recalibrating semantic features to achieve finer information aggregation. The gradient attention module (GAM) in FPUN based on regression learning is specially designed to fully fuse the gradient features so that the network can focus more on the gradient mutation regions that are improperly segmented by the PPUN. The proposed method achieved remarkable advantages over existing methods in generating datasets, synthetic fringe projection and real experiment by fully combining the advantages of semantic segmentation and regression learning.
Keywords: Measurement; Phase unwrapping; Denoise; Gradient attention module; Fringe projection