The goal of this project is to develop a Face Covering Detection (FCD) computer vision system. The system is capable of detecting whether a person is wearing a face mask correctly, incorrectly, or not wearing a mask at all. The three classes for classification are:
- Mask: The face mask is worn correctly, covering both the nose and mouth.
- No Mask: The face mask is not worn, and the face is fully visible.
- Mask Incorrect: The face mask is worn incorrectly, not covering either the nose or mouth properly.
- Real-Time Detection: Utilizes a webcam or video feed to detect and classify face masks in real-time.
- Multi-Class Classification: Distinguishes between three classes: mask, no mask, and mask incorrect.
- High Accuracy: Employs deep learning models to ensure high accuracy in detection and classification.
- User-Friendly Interface: Simple and intuitive interface for easy use and interaction.
- Cross-Platform Compatibility: Runs on various operating systems including Windows, macOS, and Linux.
- Programming Languages: Python
- Deep Learning Framework: TensorFlow/Keras and PyTorch
- Computer Vision: OpenCV
- Webcam Integration: OpenCV for capturing real-time video feed
- Pre-trained Models: Transfer learning using models like MobileNetV2 or ResNet
The dataset can be found in the link below.
In this study, four machine learning combinations were compared based on their accuracy, precision, recall, inference time, and model size:
- CNN+HOG
- SVM+HOG
- MLP+HOG
- MLP+LBP
These models were chosen due to their different strengths and weaknesses, such as size differences. HOG and LBP features were extracted from the original images to create a more condensed representation of the images.
- CNN+HOG model can be downloaded through this link: https://drive.google.com/file/d/165IkA47Tov2M3CfhYvJuQ5mCWGxkhMe1/view?usp=sharing
We tested our model on images and videos to evaluate its real-world performance. The models should be able to detect face masks in different settings, such as indoor and outdoor environments, and varying lighting conditions.
Images: We tested the models on a set of four random images, which included the 3 different classes. In most cases, the models were able to successfully identify the presence and position of face masks, even under challenging lighting conditions.
Videos: We further evaluated the performances of all four models on video files, where each model processed a continuous stream of frames in real-time. The models maintained a consistent level of accuracy throughout the video, accurately detecting face masks despite changes in subject positions, lighting, and background.
The table below shows the performances of the four models on the test set:
| Model | Accuracy | Precision | Recall | Inference Time | Model Size |
|---|---|---|---|---|---|
| CNN+HOG | 85.15% | 61.70% | 34.64% | 1.2629s | 130.18 MB |
| SVM+HOG | 82.10% | 57.43% | 56.01% | 0.2232s | 9.62 MB |
| MLP HOG | 83.62% | 79.76% | 52.86% | 0.0086s | 2.50 MB |
| MLP LBP | 83.62% | 42.15% | 40.28% | 0.0020s | 1.15 MB |
Overall, the CNN+HOG model delivered the best performance in terms of accuracy, making it the best choice for the face mask detection system. However, the MLP+LBP model can be more recommended as speed and model size are important.
[1] Wang, L., Lin, Z., Wong, A., "Covid-19 Face Mask Detection Using Deep Learning and Computer Vision Techniques," Image Processing On Line, 10 (2020), pp. 252-262. https://doi.org/10.5201/ipol.2020.288
[2] Elaziz, M. A., Hosny, K. M., Salah, A., Darwish, M. M., Lu, S., & Sahlol, A. T. (2021). Deep Learning for Face Mask Detection in the Era of COVID-19 Pandemic: