By Zhi-Song Liu, Wan-Chi Siu and Yui-am Chan

We propose a novel Conditional Variational AutoEncoder model for this Reference based Face Super-Resolution (RefSR-VAE). By using the encoder to map the reference image to the joint latent space, we can then use the decoder to sample the encoder results to super-resolve low-resolution facial images to generate super-resolution images with good visual quality.

The paper can be found in IEEE

BibTex

   @ARTICLE{8792098,  
          author={Z. {Liu} and W. {Siu} and Y. {Chan}},  
          journal={IEEE Access},   
          title={Reference Based Face Super-Resolution},   
          year={2019},  
          volume={7},  
          number={},  
          pages={129112-129126},
   }
    

Key points:

• We firstly propose a Single Image Super-Resolution via conditional Variational AutoEncoder (SISR-VAE)

• We further propose a Reference based face SR via conditional Variational AutoEncoder (RefSR-VAE) to resolve face SR with large up-scaling factors.

• Finally, we will introduce a new Reference based (RefSR-Face) dataset for the SR of face images for training and testing.

Dependencies

Python > 3.0
OpenCV library
CAFE
NVIDIA GPU + CUDA
MATLAB 6.0 or above

Complete Architecture

The complete architecture is shown as follows,

Dataset

We propose a Reference based face SR dataset. It is modified from VGGFace2. The target is to collect facial images across different sex, races and so on. Each identity should include several facial images with various poses, ages, emotions and so on. We obtain a training dataset containing 428 identities for development. Each identity includes 2~30 images. And a testing dataset also contains 428 identities. Each identity includes 1~4 images with very different appearance to the reference image. We show some examples in the following figure.

You can download the dataset from: https://connectpolyu-my.sharepoint.com/:u:/g/personal/16903300r_connect_polyu_hk/EQL52udtgg9FmqEZbkYbyh0B9fT5IPCb5z-5VyM4J8eN9g?e=2tD3YC

Implementation

You can download the pre-trained models from: https://connectpolyu-my.sharepoint.com/:f:/g/personal/16903300r_connect_polyu_hk/EvwpX5r_nEFJvRju_UqlnJ8BrU45WT2AMKBbmN7PIBkt9g?e=lHsaqG

For RefSR-VAE, run RefSR_VAE.ipynb

For SISR-VAE, run VAE-SR.ipynv

Visual Comparison

compare with state-of-the-art

This figure shows the comparison among different face SR algorithms on RefSR dataset

compare with state-of-the-art

This figure shows the facial identity transfer on RefSR dataset

Please cite our paper for using our dataset or models.