Super-Resolution Generative Adversarial Networks

Loss Function

• I just use 3 losses for G loss

1. VGG19 Bottle Neck feature loss (content loss)
2. Adversarial loss (sigmoid loss)
3. MSE loss (with generated HR img, real HR img)

• And for D loss

1. Adversarial loss (sigmoid loss) (2. maybe, change adv loss to MSE loss could be good as well i think...)

Architecture Networks

• Same as in the SRGAN paper

Tensorboard

Result

Name	Valid HR image	Global Step 1k	Global Step 40k
SRGAN

To-Do

• Results are not good as i expected... So, maybe it needs to be fixed some way soon...