Super resolution reconstruction of single image based on Res-Net and Sub-Pixel

How to improve the resolution of image is a classical problem in the field of image processing. In recent years, methods based on Convolution Neural Networks (CNN) or deep neural networks have made huge success. Researchers use a single filter behind the input layer or before the output layer to upscale Low resolution (LR) image to be a High resolution (HR) one. Moreover, too deep layers are used to capture the feature maps of the LR image, which will make the network be much more time consuming. In our work, we proposed a new approach which involves three convolution layer and two Residual Network (Res-Net) and a Sub-pixel mapping layer for the reconstruction. On the one hand, two shallow Residual Network can protect the integrity of the information of LR image to a certain extent by transferring the input information directly to the output of the next convolution layer. On the other hand Sub-pixel mapping layer which can learn complex filters to upscale the final layer of LR feature maps to be HR image. The experiments are made based on images of publicly datasets and the results as shown in this paper that our approach performs better than previous CNN-based methods, meanwhile it can perform in a real time.