Single Image Super Resolution using CNN

Abstract

Single Image super-resolution using Convolutional Neural Networks has been examined a lot. In most of the work the input used is an LR coloured image from which a lot of information has already been lost due the steps like noise reduction, compression present in the ISP(Image signal Processing) Pipeline. Here, in this work, in order to enhance the quality of the super-resolved image, another branch is added to the network. In this branch input is an unprocessed raw image obtained directly from the camera.This raw image is rich in structural details The network proposed in DCSCN[1] is treated as the baseline, which is a low complex architecture. In this, a new raw input branch is added so as to improve the quality of the generated image measured in terms of PSNR. Also, training dataset is generated using a realistic data generation pipeline introduced in RSSR[2].