GPEN | Face Restoration

GPEN

Prior embedded network (GPEN)

Model

First train a GAN prior network, and then embed it into a DNN as decoder for HQ face image restoration.

Loss

Loss = Adversaraial Loss La + Content Loss Lc + Feature Matching Loss Lf

LA is inherited from the GAN prior network

LC is L1-norm distance between the generated results and the corresponding ground-truth images.

LF is similar to the perceptual loss but it is based on the discriminator rather
than the pre-trained VGG network to fit our task.

Experiments

FFHQ，训练步骤

1. 先训练GAN网络
2. 利用合成的LQ图像去fine-tune整个GPEN网络