БАГАТОМАСШТАБНИЙ МЕТОД НА ОСНОВІ ЧАСОВОЇ ГЕНЕРАТИВНОЇ МЕРЕЖІ ДЛЯ ВИСОКОЇ РОЗДІЛЬНОСТІ ТА СТАБІЛЬНОГО РУХУ ВІДЕО

Context. The problem of improving the quality of video images is relevant in many areas, including video analytics, film production, telemedicine and surveillance systems. Traditional video processing methods often lead to loss of details, blurring and artifacts, especially when working with fast movements. The use of generative neural networks allows you to preserve textural features and improve the consistency between frames, however, existing methods have shortcomings in maintaining temporal stability and the quality of detail restoration.Objective. The goal of the study is the process of generating and improving video images using deep generative neural networks. The purpose of the work is to develop and study MST-GAN (Multi-Scale Temporal GAN), which allows you to preserve both spatial and temporal consistency of the video, using multi-scale feature alignment, optical flow regularization and a temporal discriminator.Method. A new method based on the GAN architecture is proposed, which includes: multi-scale feature alignment (MSFA), which corrects shifts between neighboring frames at different levels of detail; a residual feature boosting module to restore lost details after alignment; optical flow regularization, which minimizes sudden changes in motion and prevents artifacts; a temporal discriminator that learns to evaluate the sequence of frames, providing a consistent video without flickering and distortion.Results. An experimental study of the proposed method was conducted on a set of different data and compared with other modern analogues by the metrics SSIM, PSNR and LPIPS. As a result, values were obtained that show that the proposed method outperforms existing methods, providing better frame detail and more stable transitions between them.Conclusions. The proposed method provides improved video quality by combining detail recovery accuracy and temporal frame consistency