Zernike-informed multi-task learning framework for phase unwrapping of obstructed interferograms

Phase unwrapping is a crucial step in optical interferometric measurements, yet it becomes highly challenging when phase maps include large missing data regions caused by obstructions. This work introduces a Zernike-informed multi-task learning network (ZMMNET) that integrates Zernike polynomial coefficient regression with end-to-end phase decoding. A lightweight MobileViT backbone is employed to capture both local features and global context while the Zernike branch supplies a physics-informed representation of the wavefront. This dual-path architecture enables accurate recovery of the continuous phase across non-contiguous domains. Tests on simulated and experimental interferograms demonstrate that ZMMNET achieves higher accuracy and robustness than existing deep learning and conventional algorithms, particularly at the boundaries of obstructed regions. These results highlight its strong potential for practical applications in optical interferometry.