GENERATIVE ADVERSARIAL NETWORKS FOR IMAGE SUPER RESOLUTION

Abstract

This article demonstrates an experiment with changing one of the functions of the final image processing, the quality of which was improved using deep neural networks. It also describes previously developed methods for training and using neural networks to obtain Super-Resolution images, which have been very successful. Quantitative and qualitative assessment of the training results allows us to conclude that image processing is improved even with a very limited amount of training, and with a larger training the result can be improved even more. Within the specific goal (to get better performance or low resource consumption) and under conditions of full training, this method of image processing can be optimized for video processing. This creates the preconditions for obtaining a video quality improvement system on edge computing devices (gadgets, smartphones, tablets, etc.) and with limited computing resources and reducing the load on the network infrastructure when integrating with online video services.

Key words: convolutional neural networks, deep learning, generative adversarial networks, High-resolution, Low resolution, Super-resolution.