Image Restoration with Structured Deep Image Prior

Abstract

In this study, a novel image restoration method is proposed by introducing a structured convolutional neural network (CNN) in the deep image prior (DIP) framework. CNN has shown significance for image restoration as well as classification. DIP uses CNN structures as an image prior and shows a significant performance without explicit training of the generative model. Although CNN improves their ability by training with a large number of example images, DIP requests no training set of images for constructing the prior. The conventional DIP, however, uses a U-Net for the prior, which contains a huge number of design parameters. In order to reduce the number of parameters, this study proposes to use a structured CNN based on the non-separable oversampled lapped transform (NSOLT). By showing some simulation results, the significance of the proposed method is verified.