Children safety and privacy protection through facial characteristics censorship in photos

Final project in Deep Neural Networks course at DSIT @ UoA - Academic year 2020 - 2021

Authors:

• Alex Zerntev
• Nikolaos Episkopos
• Spyros Nikolakis

Create the dataset

Create a directory and insert some images containing faces in there. Use the full path of that directory in the PATH_TO_DATASET variable inside src/utils/face_detection_yolo.py.

$ cd src/utils
$ python3 face_detection_yolo.py

By default, the extracted faces will be stored in src/utils/faces_extracted but this can be changed by modifying the OUTPUT_PATH variable.

Finally, manually choose what faces belong to infants and what does not, the first ones should be added to a directory babies and the second ones to a directory named not-babies.

Perform predictions

In order to perform predictions with an already trained model, enter the src/predict directory. Set the PATH_TO_DEMO_IMGS and PATH_TO_EMOJIS inside the perform_predictions.py. The first directory corresponds to the directory that the images for the demo will be stored and the second one is the directory that contains the available images that will be used to hide the faces of infants. Set also the PATH_TO_MODEL variable to point to the model.h5 file of the trained model.

$ python3 perform_predictions.py

Directories Description

Train

Contains Jupyter Notebooks for model training

src/train/

       ./resnet.ipynb - ResNet - 50 Fine tuning model

       ./vggface_custom_model.ipynb - VGGFace - ResNet - 50 Fine tuning model

       ./vggface_feture_extraction.ipynb - VGGFace feture extraction and then classify via Feed Forward Neural Net

       ./visual_transformers.ipynb - Visual Transformers model training

Set the needed parameters inside of the Jupyter Notebooks, set your enviroment and the models will run out of the box.