Four-agent coherent plug & play: physics-constrained deep learning for imaging correction and wavefront sensing in deep turbulence

Using digital holographic (DH) sensors, along with iterative computational algorithms, we can sense and correct distributed-volume aberrations caused by atmospheric turbulence. These iterative algorithms combine data models accounting for the physics of rough-surface scattering, wave-propagation, and coherent detection, along with prior models to estimate both a real-valued reflectance image and a 3D wavefront. We recently developed a new framework, known as Coherent Plug and Play (CPnP), that incorporates a deep neural network as a prior model for the image. While this framework increases image quality and the Strehl ratio of our wavefront estimate, it uses a relatively simple analytical prior model for the atmospheric phase errors. In this work, we augment CPnP to incorporate deep-neural networks and transformers for both the image- and the phase-error prior models, further increasing the quality of our estimates. This improved quality allows us to push the limits of DH sensors into more challenging conditions (e.g., deep turbulence and low signal-to-noise ratio).