Noise-adaptive label-free adaptation for robust Fourier phase retrieval from experimental diffraction patterns

Fourier phase retrieval plays an important role in computational imaging and diffraction-based optical measurements. However, deep-learning-based phase retrieval methods often suffer from performance degradation when applied to noisy experimental diffraction patterns due to the domain gap between simulated training data and practical measurements. In this work, we propose a noise-adaptive label-free adaptation framework (NALA) to improve the robustness of physics-enhanced deep learning for Fourier phase retrieval. The proposed method first utilizes experimentally acquired noisy diffraction patterns for pretraining to learn realistic noise characteristics and reduce the domain gap between simulation and experiment. Based on this initialization, physics-informed fine-tuning is further performed without requiring any labels. Experiments on simulated and experimentally acquired diffraction data demonstrate that the proposed method achieves reliable object reconstruction under various noise conditions. Compared with existing approaches, the proposed framework improves computational efficiency over several baseline methods while consistently enhancing reconstruction quality in terms of PSNR, SSIM, LPIPS, and DISTS. These results indicate that the proposed approach offers an effective and practical solution for robust Fourier phase retrieval in realistic experimental environments. It is promising for diffraction imaging and experimental phase retrieval applications, where noise and measurement imperfections are unavoidable.