Address model mismatch and defocus in FZA lensless imaging via model-driven CycleGAN

Mask-based lensless imaging systems suffer from model mismatch and defocus. In this Letter, we propose a model-driven CycleGAN, MDGAN, to reconstruct objects within a long distance. MDGAN includes two translation cycles for objects and measurements respectively, each consisting of a forward propagation and a backward reconstruction module. The backward module resembles the Wiener-U-Net, and the forward module consists of the estimated image formation model of a Fresnel zone aperture camera (FZACam), followed by CNN to compensate for the model mismatch. By imposing cycle consistency, the backward module can adaptively match the actual depth-varying imaging process. We demonstrate that MDGAN based on either a simulated or calibrated imaging model produces a higher-quality image compared to existing methods. Thus, it can be applied to other mask-based systems.