Image and video super resolution using deep learning

Abstract

The aim of the Super-Resolution (SR) technique is to recover the High Resolution (HR) image from a Low Resolution (LR) image. It has applications in many computer vision areas such as medical imaging, face recognition and surveillance. Recently, many deep learning based end-to-end Super-Resolution methods have been proposed which achieves higher reconstruction accuracy than the traditional methods. Efficient Sub-Pixel Convolutional Network (ESPCN) is a CNN architecture which extracts the feature maps from the LR space and generates the HR image using the pixel shuffle method. In this project we study the effects of different settings on ESPCN model for Single Image Super-Resolution (SISR). We also extend this model using different architectures for video super resolution. We observe the performance difference on Video Super-resolution (VSR) when ESPCN architecture is modified using this various forms of spatiotemporal convolutions.