Super-resolution for noisy images via deep convolutional neural network

Super-resolution (SR) is an effective approach to enhance image spatial resolution. Although many SR algorithms have been proposed by far, little progress has been made to improve resolution for a noisy image. Conventional approaches always adopt the denoising step before applying the SR method to noisy low-resolution images. However, some high-frequency details lose during the denoising step and cannot be restored by the following SR step. Therefore, motivated by the success of deep learning in different computer vision missions, we propose a novel method named Denoising Super-Resolution Deep Convolutional Network (DSR-DCN), to combine both denoising and SR step in a single deep model. The proposed deep model straightly learns an end-to-end mapping from noisy LR space to the corresponding HR space. To equip the proposed network with the capability of blind denoising, Gaussian noise, with a range of standard deviation instead of constant value, is added to each patch of the LR space during training. Experiment results demonstrate that DSR-DCN achieves superior performance and better visual effects than the conventional approaches.