Face-Mask-detection using Multi task learning

Overview

We use multi-task learning approach for the following objectives

i) The primary task to detect faces that have their masks worn correctly or incorrectly

ii) The secondary task to detect faces that have their mask only covering the nose and mouth; masks only covering mouth and chin and mask under the mouth (i.e three cases of mask incorrectly worn)

Code Structure

This repository contains implementation of the below models for Face mask detection

1. MobileNet

    to train MobileNet for primary task - Facemask-detection-task1.ipynb
    
    to train MobileNet for secondary task - Facemask-detection-task2.ipynb
   

2. BKNet

    to train BKNet for primary task - Evalsingletask.ipynb
    
    to train BKNet for both primary and secondary tasks:
    
                Training - BKNetMultitask/BKNet_multitask_train.ipynb
                
                Evaluation - BKNetMultitask/BKNet_multitask_evaluate.ipynb
                
                Model implementation - BKNetMultitask/BKNetStyle.py
   

Code for data processing: Data Pre-processing.ipynb

Dataset

MaskedFace-Net that consists of 133,783 synthetically generated images belonging to below categories was used

1. Mask Correctly worn

2. Mask incorrectly worn

   i) Chin exposed

   ii) Nose exposed

   iii) Nose & mouth exposed

The data is accessible at https://github.com/cabani/MaskedFace-Net

Environment requirements

The code in this repo is written in Python 3. BKNet models use Tensorflow 1.13.1 and MobileNet models use Tensorflow 2.x version

Contributors

Swasthi Chittoor Shetty

Sanjana Vijay Ganesh

Samarth Varshney

Meghana Deepak

Isha Dilipkumar Shah

This work was done as part of CS6220 Big Data Systems and Analytics project requirements at Georgia Tech

References

The following papers and code were used for this project

Sang, Dinh & Bao, Cuong. (2018). Effective Deep Multi-source Multi-task Learning Frameworks for Smile Detection, Emotion Recognition and Gender Classification. Informatica. 42. 10.31449/inf.v42i3.2301. https://github.com/truongnmt/multi-task-learning

Cabani et al., "MaskedFace-Net - A dataset of correctly/incorrectly masked face images in the context of COVID-19", Smart Health, ISSN 2352-6483, Elsevier, 2020,