Answers are grounded in this paper and cite their sources.
Raw wrapped phases obtained from digital holographic speckle pattern interferometry (DHSPI) frequently exhibit high speckle noise, abrupt fringe density variations, and uneven backgrounds, posing significant challenges for reliable absolute phase retrieval. To address these challenges, we propose RPU-Net, a robust dual-decoder network comprising only 7.5 million parameters. The architecture employs a shared weighted convolution encoder to suppress background gradients, along with two decoders dedicated to absolute phase unwrapping and wrapped phase filtering, respectively. Additionally, shuffle attention and skip connections are used to restore fine details and suppress illumination disturbances. Furthermore, a wavelet-based multi-task loss function is implemented to enforce local fringe continuity and global scale consistency. Experimental results demonstrate that the proposed RPU-Net outperforms existing methods while using significantly fewer parameters. Evaluations in defect detection and various cross-domain applications reveal that RPU-Net directly produces reliable absolute phase maps without requiring post-processing or parameter tuning. Consequently, it provides a real-time, plug-and-play solution for phase unwrapping under complex background conditions.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: