PatchDIP Exploiting Patch Redundancy in Deep Image Prior for Denoising

Abstract

The structure of a deep convolutional neural network initialized with random weights is able to sufficiently capture the patterns in a natural image. This finding motivates using deep neural network as an effective prior for natural images. In this work, we show that this strong prior, enforced by the structure of a ConvNet, can be augmented with the information that recurs in different patches of a natural image to boost the performance. We demonstrate that the self-similarity in the image patches can be exploited alongside deep image prior by optimizing the network weights to fit patches extracted from a single noisy image. Our results indicate that employing deep image prior on noisy patches provides an additional disincentive for the network to fit noise, and is encouraged to exploit redundancies among the patches yielding better denoising performance.