Blind single-shot phase retrieval based on a self-supervised physics-adaptive neural network

Recently, single-shot phase retrieval techniques, which aim to reconstruct an original sample from a single near-field diffraction pattern, have garnered significant attention. Despite their promise, existing methods are highly dependent on precise physical forward models, constraining their effectiveness in real-world scenarios. To overcome the challenges posed by unknown diffraction distances in blind single-shot phase retrieval, this paper introduces a self-supervised physics-adaptive neural network termed BlindPR-SSPANN. The proposed method jointly optimizes the physical parameters of the forward propagation model alongside the trainable parameters of the reconstruction network. To achieve this, BlindPR-SSPANN incorporates a novel network architecture that integrates tunable physical parameters within a multi-stage, coupled reconstruction process. The proposed network is trained under a self-supervised scheme facilitated by a refined physics-consistent loss function. Simulation and experimental results demonstrate that BlindPR-SSPANN delivers high-performance reconstructions from a single intensity measurement, even under large diffraction distance errors, enabling self-calibrated snapshot coherent diffraction imaging.