Masks play a crucial role in protecting the health of individuals against respiratory diseases, as is one of the few precautions available for COVID-19 in the absence of immunization. Developing a real-time facial mask detection model may prove beneficial in identifying individuals who have worn masks amid the pandemic.
- Create custom datasets, dataloaders and transformers for object detection tasks.
- Build object detection models to determine the boundary boxes of human faces, and to classify them into different classes. Implement performance metrics to keep tracking the training process.
- Apply the model on images, videos and live streams to accomplish real-time face mask detection.
All the training data used in this project are from Face Mask Detection. This dataset contains 853 images belonging to the 3 classes (With mask, Without mask, and Mask worn incorrectly), as well as their bounding boxes in the PASCAL VOC format.
- Dataset Setup: Build a custom dataset for object detection tasks. Take images and annotation information from the original dataset and collate them all together.
- Image Augmentation: Apply mask augmentation methods for segmentation from
albumentations
(e.g. ShiftScaleRotate, RandomBrightnessContrast) on the train/val dataset. - Fine-Tuning Models: Build the FastRCNN model for multi-class object detection. Implement custom mean IoU score functions.
- Evaluation: Evaluate the models on the test set and track the performance of the model.
- Image Output:
- Video Output:
[1] Girshick, R. (2015, September 27). Fast R-CNN. arXiv.org. Retrieved March 20, 2023, from
https://arxiv.org/abs/1504.08083
[2] Larxel. (2020, May 22). Face mask detection. Kaggle. Retrieved March 20, 2023, from
https://www.kaggle.com/datasets/andrewmvd/face-mask-detection
[3] Techzizou. (2023, January 9). Train a custom Yolov4 object detector (using Google Colab). Medium. Retrieved March 20, 2023, from
https://medium.com/analytics-vidhya/train-a-custom-yolov4-object-detector-using-google-colab-61a659d4868
This repository is licensed under the Apache-2.0 License - see the LICENSE file for details.