Compressive Focal Plane Array Imager Reconstruction Using Learning Based Regularization

Abstract

In this paper, we develop a learning based regularization method for reconstructing compressive focal plane array imager (CFPAI). While many optimization algorithms employ proximal operators for regularization purposes such as total variation minimization, they are often inadequate to fully capture the likelihood of complex natural images. Recently, deep learning based approaches obtain promising results in different imaging problems, creating the possibility to use them as a regularizer in an optimization framework. Here, we utilize this approach in CFPAI obtaining spatially modulated and downsampled measurements of the incoming light intensity. We first formulate the problem of finding original high resolution image from its measurements as an optimization problem. Then, we solve the resulting problem using alternating direction method of multipliers (ADMM). In ADMM, we replace the proximal operator corresponding to the regularization function with a deep convolutional denoising network. Results show successful reconstruction performance in terms of reconstruction pSNR and visual quality even under significant noise levels.