Improving Image Super-Resolution Based on Multiscale Generative Adversarial Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional neural networks (CNNs) advance image super-resolution but struggle with blurred textures and lost high-frequency details.
• Perceptual networks often fail to reconstruct fine image textures accurately.

Purpose of the Study:

• To address limitations in current image super-resolution techniques, particularly the loss of high-frequency information and texture detail.
• To propose a novel generative adversarial network (GAN) for improved perceptual extreme super-resolution.

Main Methods:

• A generative adversarial network (GAN) employing multiscale asynchronous learning.
• Integration of a pyramid structure to incorporate high-frequency information across different scales.
• Utilization of a U-net discriminator for pixel consistency and LPIPS loss for enhanced perceptual supervision.

Main Results:

• The proposed method effectively restores detailed texture information in low-resolution images.
• Experiments on benchmark datasets (Set5, Set14, BSD100, SunHays80) validate the approach's efficacy.
• Improved reconstruction of fine image textures compared to existing methods.

Conclusions:

• The multiscale asynchronous learning GAN significantly enhances image super-resolution performance.
• The method successfully mitigates issues of blurred line structures and lack of high-frequency information.
• This approach offers a robust solution for restoring detailed textures in super-resolved images.