Confidence-guiding single image super-resolution through Gaussian Process Regression

The Gaussian Process Regression (GPR) framework offers a robust and intuitively probabilistic approach, presenting advantages over the less interpretable neural network models, particularly in the context of image super-resolution. Traditional applications of GPR in super-resolution have primarily focused on the posterior mean for generating predictions, neglecting the insightful predictive variance. This variance quantifies the confidence level of each prediction, a critical aspect for ensuring the reliability of the reconstructed images. In this study, we leverage this underutilized metric by incorporating a confidence regularization term into the iterative back projection process. This innovation aims to enhance the reconstruction fidelity by guiding the enhancement process in a more informed manner. Our extensive experimental analysis on established benchmark datasets confirms the superior performance and efficiency of our proposed approach, marking a significant advancement in super-resolution methodologies.