FRGAN: A Blind Face Restoration with Generative Adversarial Networks

Table 2

PSNR- and SSIM-based superresolution performance on 300 W balanced dataset across pose (higher is better).

Method	PSNR			SSIM
	30	60	90	30	60	90
Bicubic	21.03	20.20	20.98	0.7304	0.7597	0.7902
PFSR	20.42	21.96	21.76	0.7675	0.7871	0.7965
SR-GAN	20.01	20.94	21.48	0.7269	0.7465	0.7586
DIC	19.06	20.91	20.86	0.7462	0.7521	0.8014
Ours-i	21.49	22.53	23.28	0.7963	0.8056	0.8301
Ours-ii	21.38	22.66	23.54	0.7985	0.7948	0.8199
Ours-iii	21.57	22.35	23.32	0.7991	0.8095	0.8158
Ours-iv	20.60	21.72	22.61	0.7721	0.7964	0.8009

The results are not indicative of visual quality.