Super-Resolution for Face Recognition Improvement

Repository based on:

• LightCNN
• SRGAN

Data

If you want to reproduce work with datasets:

1. Download and place data in corresponding directories
   • Aligned LFW -> data/datasets/lfw/lfw_funneled
   • Aligned and cropped MS-Celeb-1M -> data/datasets/ms_celeb/FaceImageCroppedWithAlignment.tsv
   • Filtered list of MS-Celeb-1M -> data/datasets/ms_celeb/MS-Celeb-1M_clean_list.txt
2. And run dataset_preparing.ipynb

Weights for LightCNN -> data/weights/light_cnn

Train

Trained on MS-Celeb-1M dataset. You can reproduce train pipeline using srgan_training.ipynb

Results

All approaches tested on LWF 6000 pairs.

	ROC AUC
Real HR	0.98207
SRGAN with Light CNN 9 (MFM4) NO ADVERSARIAL	0.96832
SRGAN with Light CNN 9 (MFM4)	0.96742
SRGAN with Light CNN 9 (FC) NO ADVERSARIAL	0.96594
SRGAN with LIGHT CNN 9 (MFM4) NO ADVERSARIAL NO IMAGE	0.96421
SRGAN with VGG (3.1)	0.96349
SRGAN with VGG (3.1) NO ADVERSARIAL	0.96346
SRGAN with MSE	0.95951
Bicubic interpolation	0.93559

You can reproduce results using recognition_test.ipynb