Image demosaicing using Deep Image Prior

Abstract

The paper focuses on the problem of image demosaicingusing the deep image prior. The deep image prior (DIP)is an uncommon concept that uses a generative neural networkwhich, however, utilizes only the degraded image as the inputfor training. A novel method for image demosaicing is proposed,based on DIP, and it is compared with common demosaicingmethods. In terms of the objective PSNR and SSIM values,the proposed method proved to be comparable with a widelyused Malvar’s demosaicing method. Nevertheless, subjectively,DIP produces demosaiced images comparable with the superiorMenon’s algorithm. Unfortunately, the proposed method turnedout to be computationally immensely challenging