Image Super-Resolution Using Generative Adversarial Networks

The goal of this work is to implement an algorithm for single image super-resolution using generative adversarial networks. In this paper, practical limitations of increasing image resolution using traditional methods are considered. In addition, modern solutions to this problem, with and without the use of convolutional neural networks, are described and compared. The advantages of Generative Adversarial Networks (GANs) over traditional convolutional neural networks are discussed, as well as network’s generator modifications that allow to achieve state-of-the-art results. A comparative study of the visual quality of the reconstructed images is carried out. The developed system allows the operator to experiment with various network hyperparameter configurations. The network itself is implemented with Theano and Lasagne. The trained models are then used in a web application that demonstrates the capabilities of the implemented algorithm. This paper contains 39 pages, 3 chapters, 20 illustrations, 3 tables, 38 bibliography items, 4 appendices.