Image upscaling using generative adversarial networks

Abstract

The purpose of this article is to develop and improve image scaling algorithms aimed at preserving details and visual appearance when changing image size. The subject of this article is image scaling algorithms, particularly their enhancement to ensure high-quality images. An analysis of artifacts arising during image scaling are conducted. Additionally, various image scaling algorithms are analyzed. Special attention is given to algorithms and methods of image scaling that use machine learning and neural networks. In this article a software implementation of the image scaling method using Generative Adversarial Networks is proposed. The architecture of GAN is described in detail and the loss functions for generator and discriminator are calculated. The results of this article demonstrate that improved algorithms based on GANs allow achieving high-quality image scaling with minimal detail loss and artifacts. The software implementation of the algorithms showed efficiency and the possibility of real-time image scaling, which is crucial for various practical applications such as medical imaging, video processing, and other fields where image quality is critical.