nepiskopos / children_privacy_protection

Machine learning pipeline for hiding small children faces in pictures to conceal their identity and protect their privacy.

Children safety and privacy protection through facial characteristics censorship in photos

Final project in Deep Neural Networks course -- academic year 2020-2021

MSc in Data Science and Information Technologies, NKUA

Authors:

• Alex Zerntev
• Nikolaos Episkopos
• Spyros Nikolakis

Description

In this project we attempt to create a facial censorship system which hides the faces of young children in images by replacing them with emojis. To this purpose we employ:

• YOLOFace for face detection
• vit-keras for face classification (child / non-child)

Dataset

The dataset we used was a custom dataset which consisted of images of people. These images were mainly harvested from the web. We also used some pictures of our families and relatives. Because of this, the dataset will NOT be shared in this repository or anywhere else.

Project structure

Part 1: Face detection and extraction

Contains 1 Jupyter Notebooks

The corresponding notebook face_detection_extraction.ipynb performs face extraction from the original images in our dataset.

Afterwards, the extracted faces need to be manually classified and labelled as child / non-child and afterwards split into 3 distinct datasets:

• training set
• validation set
• test set

Part 2: Model training and evaluation

Contains 8 Jupyter Notebooks

The corresponding notebooks perform model training and evaluation as well as model comparison.

Notebook name	Description
evaluation.ipynb	Evaluate and compare all trained models
resnet_fine_tune.ipynb	Fine-tune a ResNet-50 model on our data
resnet_train.ipynb	Train a ResNet-50 model on our data
vggface_resnet_fine_tune.ipynb	Fine-tune a VGGFace (ResNet-50) model on our data
vggface_resnet_train.ipynb	Train a VGGFace (ResNet-50) model on our data
vggface_vgg_fine_tune.ipynb	Fine-tune a VGGFace (VGG16) model on our data
vggface_vgg_train.ipynb	Train a VGGFace (VGG16) model on our data
vision_transformers.ipynb	Train a Vision Transformer model on our data

Part 3: Classification and face censorship

Contains 1 Jupyter Notebook

The corresponding notebook prediction_censorship.ipynb performs face detection, class prediction and children's face censorship on a subset of the original images in our dataset.

Instructions

First of all, you need to use some Linux distribution and have git installed on your OS.

Set the needed parameters (paths) in every Jupyter notebook, set up your enviroment and the everything will run out of the box.

Follow the parts order when running any notebooks.

Create the dataset

1. Create a directory named dataset/.
2. Under dataset directory create a subdirectory named original/.
3. Place some images containing faces inside original.
4. Set the path to dataset/original/ in the DATASET_PATH variable both in face_detection_extraction.ipynb and prediction_censorship.ipynb.
5. You are good to go.

By default, the extracted faces will be stored under dataset/faces_extracted/ but this can be changed by modifying the FACES_PATH variable in face_detection_extraction.ipynb.

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 and face censorship with an already trained model, you need to have a directory with some emoji icons under dataset/ e.g. dataset/emojis/.

In face_detection_extraction.ipynb set the DATASET_PATH, EMOJIS_PATH, MODEL_PATH and SAVE_PATH to match the dataset path, the emoji directory path, the trained model path and the output directory path.

A random sample of the original images will be selected for the demo.

The results will be placed in the directory selected in SAVE_PATH.