Research on Generative adversarial networks for Super Resolution

Reference papers: :bookmark_tabs:

• Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
• ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks
• Generative Adversarial Network for Image Super-Resolution Combining Texture Loss
• And some related papers

Environment and platform 💻

• Tensorflow 2.7
• Google Colab

The project is implemented on Google driver.

---

Result

Implementation

Architecture:

• Generator

---

• Discriminator

Losses:

I implemented 3 loss terms to support Generator reconstruct Super Resolution image:

• Pixel-wise Loss
  
• Feature Loss
  
• Style Loss
  

---

And with Adversarial Loss I used LSGAN combine with Relativistic average GAN:

• For Generator
  
• For Discriminator
  

Training