Video resolution enhancement using deep neural networks and intensity based registrations

Abstract

Thanks to the recent rapid improvements made to the maximum possible resolution of display devices, higher qualities of experience have been made possible, which necessitates either producing and transmitting considerably higher volumes of data or super-resolving lower-resolution contents at the display side, where the former might not be practically feasible. Therefore, aiming at the latter, this paper proposes a novel super-resolution technique, which takes advantage of convolutional neural networks. Each image is registered into a window consisting of two frames, the second one standing for the reference image, using various intensity-based techniques, which have been tested and compared throughout the paper. According to the experimental results, the proposed method leads to substantial enhancements in the quality of the super-resolved images, compared with the state-of-the-art techniques introduced within the existing literature. On the Akiyo video sequence, on average, the result possesses 5.38dB higher PSNR values than those of the Vandewalle registration technique, with structure adaptive normalised convolution being utilized as the reconstruction approach.