The talk will introduce a problem of upscaling a picture with as small loss of quality as possible using deep learning techniques. What metric to use when evaluating the solution? Upsample the image in the beginning or near the end of your neural network? Which upscaling layers to use? Answers to these and more questions on this topic will be discussed.
Single Image Super Resolution is a problem where, having a single low-resolution image, one wants to produce an estimate of the corresponding high‑resolution image. Several neural network architectures that tackle this problem will be described and compared. Special attention will be paid to different upscaling layers in neural networks and their properties. Also, the concept of perceptual loss and its application to Super Resolution will be presented. All of these topics will be supplemented with relevant experience from developing research code in PyTorch for this problem. Finally, results of the research will be presented.