부산대학교 도서관

This research proposes an approach to enhance the denoising and upscaling performance of noisy images using Generative Adversarial Networks (GANs), particularly Style GAN architecture. Denoising and upscaling noisy images are crucial in many computer vision applications, and GANs have shown remarkable effectiveness in creating high-quality images. However, training Style GAN requires huge amount of data and is computationally expensive. To address this issue, this study proposes using various filters such as mean, median, and weighted median to pre-process the noisy images before feeding them to Style GAN. The proposed approach achieves superior denoising and upscaling compared with other system in terms of FID and inception score, and further exploration of hyperparameters and variations of the Style GAN architecture can lead to even better results.