Relative adversarial learning for photo-realistic single image super-resolution
Zhang, Meilin (2025)
Kandidaatintyö
2025
School of Engineering Science, Tietotekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2025051544968
https://urn.fi/URN:NBN:fi-fe2025051544968
Tiivistelmä
The application of generative adversarial networks (GANs) in single image super-resolution has led to remarkable improvements in perceived image quality, they achieve better visual fidelity compared to conventional interpolation and convolutional neural network approaches. Although GAN-based methods often perform worse on objective metrics such as PSNR, they achieve better results in terms of visual perception, highlighting the gap between pixel-level accuracy and perceptual quality.
However, GANs often suffer from unstable training and convergence issues, which limit their practical application. To address this problem, this thesis proposes a relativistic generative adversarial network (RaGAN). By using a relativistic discriminator, the model learns relative realness instead of absolute realness, which helps stabilize training and improves the texture and realism of generated images.
To evaluate performance, the method was tested on three standard datasets used in super-resolution research: Set5, Set14, and BSD100. Experimental outcomes indicate that while the method exhibits marginally inferior performance on quantitative evaluations (PSNR, SSIM) compared to conventional approaches, it demonstrates notable advantages in percep -tual fidelity. This observation aligns with existing findings that improvements in established PSNR may not consistently correlate with enhanced visual realism.
However, GANs often suffer from unstable training and convergence issues, which limit their practical application. To address this problem, this thesis proposes a relativistic generative adversarial network (RaGAN). By using a relativistic discriminator, the model learns relative realness instead of absolute realness, which helps stabilize training and improves the texture and realism of generated images.
To evaluate performance, the method was tested on three standard datasets used in super-resolution research: Set5, Set14, and BSD100. Experimental outcomes indicate that while the method exhibits marginally inferior performance on quantitative evaluations (PSNR, SSIM) compared to conventional approaches, it demonstrates notable advantages in percep -tual fidelity. This observation aligns with existing findings that improvements in established PSNR may not consistently correlate with enhanced visual realism.