Single Image Super Resolution Using Convolutional Neural Networks for Noisy Images

In this paper, we address a problem of image super resolution to obtain a noise-free and high resolution image from a noisy and low resolution image. In recent years, deep learning-based approaches have been achieved a lot of progress to the image restoration problems. However, it is still not trivial to generate a high quality image when the input image is both noisy and low-resolution, because it is difficult to disambiguate the fine texture components from noise components for the input image. In this case, conventional super-resolution algorithms usually amplify the noise along with the details. To deal with this problem, we propose a super-resolution network that is robust to noisy images by constructing multi-modules in parallel architecture. The experimental results show that our proposed network restores a noise-free and rich-texture image from the low-resolution and noisy input image, while other methods fail.

I. Introduction

The image super resolution (SR) is among popular image restoration problems that have received much attention for decades. The goal of SR is to restore a high resolution (HR) image from a low resolution (LR) image by estimating texture details that were eliminated during down-sampling process. Recently, extensive studies based on deep learning have been carried out on image SR [1]–[6].