¹ Phase Demodulation of Single-Frame Interferograms Based on MultiTask and Transfer Learning

Deep learning-based phase demodulation of single-frame interferograms offers a promising solution for high-speed, high-precision optical measurements. However, the large and fluctuating range of the unwrapped phase, coupled with significant domain gaps between simulated and experimental interferograms, poses challenges to network generalization and practical deployment. To address this, we propose a phase demodulation framework that integrates a collaborative multi-task network with a tailored transfer learning strategy. The proposed network adopts a multi-task design that decouples phase demodulation into two complementary branches: a classification branch that captures the periodic structure through wrapped count prediction, and a regression branch that focuses on intra-period details by estimating the wrapped phase. To generate the final unwrapped phase map, the outputs of both branches are fused by the decision fusion module. To bridge the domain gap between simulated and experimental data, a transfer learning pipeline is introduced: the network is first pretrained on a large-scale simulated dataset and then fine-tuned on a limited number of experimental samples. During fine-tuning, the backbone is frozen, and only the multi-task and fusion modules are updated, ensuring efficient adaptation with minimal data. Experimental results demonstrate that the proposed method achieves an RMS error of 0.0068λ and a PV error of 0.0977λ on real interferograms, validating its effectiveness for robust, end-to-end phase demodulation in practical scenarios.