International Joint Conference on Artificial Intelligence 2017

Learning to Hallucinate Face Images via Component Generation and Enhancement

Yibing Song1     Jiawei Zhang1     Shengfeng He2     Linchao Bao3     Qingxiong Yang4    

Topic Background

Face hallucination aims to generate high resolution face images from low resolution inputs. It has a series of applications ranging from video surveillance to image editing. Under the data-driven framework which consists of low and high resolution image pairs, the main difficulty is to establish the correspondence between the low resolution input and high resolution training data. In comparison, existing GAN based CNN networks contain the limitation that they do not favor high numerial scores. These networks are not precisely set to measure the pixel wise difference. It results in the incorrect alignment between the upsampled facial structure output and that of ground truth even though the overall perception seems more appropriate.


Prior art methods : Low resolution correspondence generation + high resolution texture transfer + post processing refinement.
Proposed : Deep facial component generation + high resolution structure synthesis + detail transfer.

Contributions: State-of-the-art performance under scaling factor of 4 and 10.



[Paper.pdf] : The paper.
[] : Complete results on the benchmark datasets.
[] : A self-contained C++ (VS2012) / matlab implementation. Training dataset is also included.
[Slides.pptx] : The slides.
[Poster.pdf] : The poster.

BibTex (DOI)

  author    = {Song, Yibing and Zhang, Jiawei and He, Shengfeng and Bao, Linchao and Yang, Qingxiong},
  title     = {Learning to Hallucinate Face Images via Component Generation and Enhancement},
  booktitle = {International Joint Conference on Artificial Intelligence},
  pages     = {4537--4543},
  year      = {2017},