Optically Coherent Imaging using Four-Agent Plug-and-Play

Optically-coherent imaging offers significant advantages in resolution and sensitivity compared to non-coherent sensors. However, the quality of the resulting images is often compromised by issues such as speckle, noise, and phase errors caused by atmospheric turbulence. State-of-the-art image reconstruction frameworks for coherent systems use a plug-and-play framework that couples physics-based models with image models learned by a convolutional neural network (CNN). These existing frameworks focus primarily on image reconstruction. In contrast, we develop a new algorithm that extends the plug-and-play approach by incorporating four agents to solve for both the image and the atmospheric phase errors. Two of these agents employ CNNs for the image and phase-error models, replacing the analytical models previously used. The CNNs implicitly learn the prior models for both the image and phase errors by leaning to remove synthetic noise added to a training set of clean images. In this work, we evaluate the performance of our new algorithm in simulation. Our results demonstrate a significant improvement in image quality.