Solving Inverse Computational Imaging Problems using Deep Pixel-level Prior

Generative models based on deep neural networks are quite powerful in modelling natural image statistics. In particular, deep auto-regressive models provide state of the art performance, in terms of log likelihood scores, by modelling tractable densities over the image manifold. In this work, we employ a learned deep auto-regressive model as data prior for solving different inverse problems in computational imaging. We demonstrate how our approach can reconstruct images which have better pixel-level consistencies, as compared to the existing deep auto-encoder based approaches. We also show how randomly dropping the update of some pixels in every iteration helps in a better image reconstruction. We test our approach on three computational imaging setups: Single Pixel Camera, LiSens and FlatCam with real and simulated measurements. We obtain better reconstructions than the state-of-the-art methods for these problems, in terms of both perceptual quality and quantitative metrics such as PSNR and SSIM.