Adaptivni Sistemi Avtomatičnogo Upravlinnâ (May 2023)
Image generations techniques using Generative adversarial networks
Abstract
The object of research is image generation algorithms based on GAN. The article reviews the main uses of these networks for image generation and main types of such algorithms, which can be used for this. Generative Adversarial Networks (GANs) have been a significant breakthrough in machine learning, allowing the generation of images that are indistinguishable from those created by humans. Although GANs have only been around since 2014, there get significant improvements due to changes in algorithms, usage of bigger datasets, and increase of computing power over the past eight years, resulting in various modifications of the network that are actively used today. Generally, all GANs can be divided into four main categories: Conditional GAN (CGAN), Progressive GAN (PGAN), StyleGAN, and CycleGAN, which are used for different tasks and cover most of the use cases of described algorithm. The GAN model consists of two main parts: a generator and a discriminator. The generator creates new instances from input data in the latent space, while the discriminator determines whether the instances are real or fake. Both models are trained based on the predictions of the discriminator, while coefficients are changed based on the MinMax algorithm. After that, some of the main modifications, such as StyleSwin, CWGAN, Layered Recursive GAN and CVAE-GAN were described, They can be used to improve the model and its main parameters such as the learning speed of the model, the quality of the obtained result and the number of artifacts that can appear during its operation. Ref. 13, pic. 5
Keywords