VIDEO UPSAMPLING OF CCTV FOOTAGES

Abstract

The idea of super resolution and image upsampling have taken the field of computer vision by storm. New methods to upsample a grainy and low resolution videos are now the new chase. Our research is focused on upsampling a CCTV video through the use of deep learning techniques. Video superresolution often show sub-par results because they tend to have more components to process than their image counterparts, namely temporal dimension apart from the usual spatial dimension. In this research, we have studied these components and developed a pipeline that effectively processes the spatio-temporal information through optical flow, backed up by novel deep learning based VSR practices such as feature alignment, aggregation and upsampling. We examined and improved the pipeline based on the BasicVSR architecture and developed a model of our own by introducing residual in residual dense blocks. The new model RD-BasicVSR, was successful in surpassing the results of BasicVSR in both PSNR and SSIM metrics at same experimental settings.