In this project, a simple tensorflow implemented autoencoder is trained on 50 young and 50 old faces to create a face aging application.
UTKFace dataset is used for this exercise, particularly the repository of aligned and cropped faces.
The images in this dataset are sorted into two folder "data/utk/young" and "data/utk/old". The code for this process is written in preprocess-utkcropped.py
Two architectures of autoencoder were trained using tensorflow library :
Data preprocessing steps such as resizing and normalization are performed on images in both young and old folders before passing it into the autoencoder network.
The results of the VGG16 encoder and decoder network on UTKFace dataset are in Results-Autoencoder-VGG.ipynb notebook.
Training with a larger sample and more epochs can potentially result in better face aging capability of the autoencoder.