Dr. Heuna Kim Mathematics and Computer Science

Paper Review for Photo-Realistic Single Image Super-Resolution (SISR) Using a GAN


Main Contribution

  • SRResNet : set a new best benchmark result for SISR for PSNR measure
  • SRGAN:
    • Content loss: a variant of pixel-wise loss depending on the network feature map
    • Adversarial loss: the probability of the discriminator over all training samples \[ l_{Gen}^{SR} = \sum_{n=1}^N - log D_{\theta_D}(G_{\theta_G}(I^{LR})) \]
    • GAN based minmax between generated and original \[ \min_{\theta_G} \max_{\theta_D} E_{I^{HR} \sim p_{train}(I^{HR})} [\log D_{\theta_D}(I^{HR})] + E_{I^{LR} \sim p_G(I^{LR})} [\log(1- D_{\theta_D}(G_{\theta_G}(I^{HR})))] \]
    • performance improvement on MOS testing by a far margin

Relevant Terminologies to understand

  • perceptual similarity
  • SSIM - structural similarity
  • PSNR - peak signal-to-noise ratio
  • MOS - mean opinion score
  • Wilcoxon signed-rank tests
  • Parametric ReLu

Interesting Relevant Work