Noise robust single image super-resolution using a multiscale image pyramid

Abstract Self-learning-based single image super-resolution (SLSR) generates a high-resolution (HR) image by estimating a mapping function between patches of different resolutions. However, when handling noisy images, SLSR methods often incur a severe noise amplification problem, or suffer from image quality degradation from additional denoising procedure since noise correlation is pre-reinforced in the super-resolving process. To achieve super-resolution while preserving image quality, we propose a unified framework of super resolution and noise reduction by taking advantage of a multiscale image pyramid generated from iteratively blurring and subsampling input and output images. A salient feature of the proposed method lies in its more elaborate mapping on the basis of high-order derivatives that are readily available from the self-similarity property of the image pyramid. The proposed method is further augmented by incorporating a hybrid interpolation and denoising scheme that leverages noise decreasing property inherent in the image pyramid, thereby minimizing noise amplification in conventional SLSR. Our experiments on both synthesized and real life images demonstrate that the proposed method yields visually more appealing HR images than the conventional sequential pipeline, and improves quantitative metrics for image quality such as the peak signal-to-noise ratio and structural similarity as well.