2. Motivation
Problem:
To design a model that renders natural face aging without image
pairs.
Applications:
Automatic face aging for cross-age face recognition.
Now + 20 years- 20 years
3. Scope
To learn a representation of face aging.
To generate artificial images of face aging for a
set age groups.
4. Image Data
Properties:
- Face image with known age.
- Age distribution.
Source:
IMDB-Wiki data set with celebrities’ images:
500k+
Experiments:
- Compared different level pre-processing.
- Compared two image sizes.
8. Approach: AC-GAN
Semi-supervised learning GANs:
- Use deep convolutional GANs (DC-GAN).
- Add class labels for images.
- Auxiliary contrained GAN (AC-GAN)
Experiments:
Compare different parameters:
13. Summary
Summary:
GAN model learned a representation of face aging.
Challenges:
- Image quality depends on pre-processing and data cleaning.
- Model tuning is challenging.
Blog post:
www.github.com/mbhuber/Insight-AI-Project