Unsupervised Deep Image Prior-Based Neural Networks for Single Image Super-Resolution: Comparative Analysis and Modelling Guidelines

Abstract

Deep Image Prior (DIP) has been recently introduced as a method to exploit the structural priors inherent to neural networks. In the field of image processing, DIP effectively addresses various problems such as denoising, inpainting, image restoration and super-resolution. Unlike supervised neural networks, which require large amounts of labelled data, DIP operates as a single-image method, where prior knowledge is derived directly from the architecture of the neural network. In this work, we focus on the single-image super-resolution problem using DIP. Through extensive experiments for image super-resolution, we show that the original formulation of DIP can be improved by properly modelling fidelity with multiple down-sampling operators. Our experimental results systematically explore combinations of regularisation and fidelity terms across both hyperspectral and natural RGB image datasets, offering new guidelines for developing effective DIP-based approaches. Code and data are available at https://github.com/capo-urjc/dip-sisr.