Image Super-Resolution Based on Variational Autoencoder and Channel Attention
Jian Xu, Yurong Zhao · 2023
Super-resolution (SR) method based on generative adversarial networks (GANs) has achieved excellent performance in both visual perception and image quality. However, there is still room for improvement. Therefore, we propose a variational autoencoder (VAE) network architecture. The VAE encoder can learn the probability distribution of the low-resolution (LR) image and reflect the probability with a latent variable, and the decoder restores the original image through latent variables. The VAE and discriminator work together to effectively distinguish between generated images and real high-resolution (HR) images. In addition, we introduce a channel attention (CA) mechanism into the discriminator to improve the cohesion between channels and extract useful features more effectively. With the help of VAE and CA, the proposed method achieves not only higher peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) values, but also more realistic visual quality. The experimental results verify the feasibility of the proposed method.