Rethinking deep image prior for denoising

Abstract

Deep image prior (DIP) has been groundbreaking for diverse inverse problems with no training required. Among the problems, however, denoising is known to be particularly chal- lenging for the DIP. We significantly improve the DIP for denoising with extended Stein’s unbiased risk estimator (eSURE) and a loss based stopping criterion that leads to a far better solution without early stopping under Gaussian noise, and Poisson noise as an extension. We analyze our method by the notion of effective degrees of freedom. Given a single noisy image, our method denoises it while preserving rich textual details by outperforming state- of-the-art denoising methods in LPIPS by large margins with comparable PSNR and SSIM in extensive experiments.