Research Article

FRGAN: A Blind Face Restoration with Generative Adversarial Networks

Table 1

Performance comparison on AFLW2000-3D and AFLW-LFPA.

MethodsAFLW2000-3DAFLW-LFPA
0° to 30°30° to 60°60° to 90°MeanMean

SDM3.674.949.676.12
3DDFA3.784.547.935.42
3DDFA + SDM3.434.247.175.42
PAWF4.72
Yu et al.3.626.069.56
3DSTN3.154.335.984.49
DeFA4.503.86
PRNnet2.753.514.613.622.93
2DASL2.753.444.413.532.85
Ours2.713.424.383.442.84

The NME (%) for faces with different yaw angles is reported. The numbers in bold are the best results on each dataset; that is, the lower is the better. “—” indicates that the corresponding result is unavailable.