SR-GAN
Today I learned about SR(Super Resolution)-GAN, so I’d like to share it here.
In previous research, super-resolution tasks(Enhancing resolution) struggled when recovering finer text details at large upscale factors. SR-GAN is the first to be able to infer images for 4x upscaling factors.
Adversarial Network Architecture
The high-resolution images are only available during training, so we generate the corresponding low-resolution image by applying a Gaussian filter and down-scale by factor r. The generator will use this low-resolution image to restore back to the high-resolution image. The discriminator will try to distinguish whether the image is a restored image or not.
Perceptual Loss
By combining adversarial loss and content loss we get Perceptual loss. For content loss, the paper uses a VGG loss (euclidean distance between the feature representations of a reconstructed image) instead of MSE loss(pixel-wise). Unlike MSE loss which is motivated by similarity in pixel space, VGG loss is motivated by perceptual similarity.
