Phase retrieval algorithms are now an important component of many modern computational imaging systems. A recently proposed scheme called generalized expectation consistent signal recovery (GEC-SR) shows better accuracy, speed, and robustness than numerous existing methods. Decentralized GEC-SR (deGEC-SR) addresses the scalability issue in high-resolution images. However, the convergence speed and stability of these algorithms heavily rely on the settings of several handcrafted tuning factors with inefficient turning process. In this work, we propose deGEC-SR-Net by unfolding the iterative deGEC-SR algorithm into a learning network architecture with trainable parameters. The parameters of deGEC-SR-Net are determined by data-driven training. Numerical results show that deGEC-SR-Net provides substantially faster convergence than deGEC-SR and exhibits superior robustness to noise and prior mis-specifications.

中文翻译：

---

学习展开期望一致信号恢复算法的相位检索


相位检索算法现在是许多现代计算成像系统的重要组成部分。最近提出的一种称为广义期望一致信号恢复（GEC-SR）的方案比许多现有方法表现出更好的准确性、速度和鲁棒性。去中心化 GEC-SR (deGEC-SR) 解决了高分辨率图像的可扩展性问题。然而，这些算法的收敛速度和稳定性严重依赖于几个手工调整因子的设置，转动过程效率低下。在这项工作中，我们通过将迭代 deGEC-SR 算法展开为具有可训练参数的学习网络架构来提出 deGEC-SR-Net。 deGEC-SR-Net 的参数由数据驱动的训练确定。数值结果表明，deGEC-SR-Net 的收敛速度比 deGEC-SR 快得多，并且对噪声和先前的错误规格表现出卓越的鲁棒性。