Abstract
Generating models are a modern, rapidly developing direction of research in the application of neural network technologies. Its currently known applications include, in particular, image visualization, drawing, image resolution improvement, structured prediction, research in the field of training and data preprocessing for neural networks in cases where the production of test data is an unacceptably expensive process. They are actively discussing their capabilities in the tasks of providing information and computer security, tasks of computer technical expertise, tasks of transforming and restoring objects by fragments or when the information is extremely noisy.
Generating models are a modern rapidly developing field of research in neural network technologies. Its applications include image visualization, drawing, image resolution improvement, structured prediction, research in the field of providing information security, tasks of computer technical expertise, tasks of transforming and restoring fragmented or is extremely noisy objects. The aim of the study is to create a software product that tests the hypothesis of the successful use of generative adversarial networks (GAN) to improve the characteristics of images. The use of GAN allows solving the problem of correcting distorted digital objects not by traditional adjustment of their individual parameters, but by generating undistorted objects that are “indistinguishable” from the original ones.
The article presents and analyzes the method of building a neural network that is capable to remove noise from images, preventing them from being too blurred and retaining clarity (in comparison with the original), which proves its ability to generalize. The experiment showed that GAN with sufficient efficiency not only removes noise from images with a distorting signal on which it was trained, but also in the case of noise unknown to the etymology network. The success of training to suppress specific distortions of digital graphic objects makes it possible to assume that by training a neural network to work with other types of distortion, one can achieve successful results for solving a wider class of problems.