Model-based network architecture for image reconstruction in lensless imaging

We introduce a multi-branch model-based architecture for image reconstruction in lensless imaging. The structure consists of two learning branches, namely a physical model-based network, and a data-driven network. It uses intermediate outputs from the former as a prior for guiding the learning of the reconstruction neural network, which mimics the mapping between the reconstructed high-resolution images and raw images. We demonstrate that the proposed architecture offers a flexible combination of model-based methods and deep networks with superior reconstruction performance than methods using only an unrolled optimization network or pure deep neural networks for image reconstruction.