Face Mask Detection using PyTorch

A simple Face Mask Detection Model created using PyTorch and OpenCV.

📃 Installation

1. Clone the repo

$ git clone https://github.com/rakshit087/facemaskDetectorPytorch.git

2. Now, cd the cloned repo and install the dependencies by (You can use virtualenv for this purpose)

cd /YOUR_PATH/facemaskDetectorPytorch
$ pip3 install -r requirements.txt

💻 Usage

🔹The faceExtractor can be used to extract face images from pictures. This is what I used to improve my dataset. Put your Images in the 'Images' folder and all the Extracted faces will saved in the 'saved' folder after running the app.py

$cd /YOUR_PATH/facemaskDetectorPytorch/Face_Extractor 
$python app.py

🔹You can train your own model, but for that first you need your own dataset. The dataset must have the following structure-

.Train_Model
├── data                       # your dataset folder
│   ├── train                  # to train your data
│       ├── mask
│       └── no_mask
│   └── test                   # to test your model
│       ├── mask
│       └── no_mask
└── Train_Model.py 

After creating the dataset you can train your model by using 'Train_Model.py', after training the model, if you are satisfied by the results you can save the model when asked. The trained model will be saved in the same directory.

$cd /YOUR_PATH/facemaskDetectorPytorch/Train_Model 
$python Train_Model.py

🔹 You can use a pretrained model with the help of 'faceDetect.py', by default it will load the model I trained.

$cd /YOUR_PATH/facemaskDetectorPytorch/Use_Model 
$python faceDetect.py

📁 Dataset Used

I created a custom dataset by mixing images from various sources.
🔹Cabani's Dataset - I used it to get images of people wearing mask
🔹Real vs Fake Face - I used it to get images of people not wearing a mask.
🔹I also some some personal images of my friends and family and extracted faces with the help of 'Face_Extractor'

🧠 Model Details

The model used Transfer Learning using MobileNetv2 (The parameters were freezed) as the base and I changed the classifier to -

Linear Layer (input - 1280 | output - 256 | Activation - ReLU)

⬇️

Linear Layer (input - 256 | output - 128 | Activation - ReLU | Dropout = 0.4)

⬇️

Linear Layer (input - 128 | output - 64 | Activation - ReLU)

⬇️

Linear Layer (input - 64 | output - 32 | Activation - ReLU | Dropout = 0.4)

⬇️

Linear Layer (input - 32 | output - 2 | Activation - SoftMax)

⚠️ Issues and Limitations

🔹 The model is having a hard time to detect dark masks, I tried to improve it by adding some dark mask images but somhow, the model started giving false positive to by beard face. 😛

💜 Thank You

Resources, study material and everything in between -

Deep Learning Specialization I learned deep learning from here, after completing the 4th Course, I decided to start this project.

PyTorch Crash Couse - By Sentdex. I learned Pytorch from here. Really awesome playlist with to the point explanation.

MobileNet with Pytorch I used official documentations as a reference for transfer learning.

Chandrika Deb's Mask Detector It is a similar project, so whenever I was lost, I used this as a reference. Also the Readme of this project is heavily inspired by her's

PS-I recently learned deep learning, so there may be a better approch to this project. Any kind of improvement or suggestions are most welcomed.