Face mask detection and classification via deep transfer learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Public Health

Background:

• Mask-wearing is crucial for preventing COVID-19, reducing infection rates by up to 40%.
• Real-world mask detection faces challenges like poor lighting, occlusion, and varied mask types (cotton, sponge, scarves), diminishing protection.
• Existing detection methods struggle with accuracy and classifying mask quality.

Purpose of the Study:

• To develop an advanced algorithm for accurate face mask detection and classification.
• To improve the reliability of mask detection in diverse real-world scenarios.
• To differentiate between qualified (e.g., N95, medical) and unqualified (e.g., cotton, scarves) masks.

Main Methods:

• Proposed a novel face mask detection algorithm integrating transfer learning with Efficient-Yolov3, utilizing EfficientNet as the backbone and CIoU loss.
• Developed a mask classification algorithm combining transfer learning and MobileNet, trained on a custom dataset of qualified and unqualified masks.
• Created a mask classification dataset categorizing masks into qualified (N95, disposable medical) and unqualified (cotton, sponge, scarves).

Main Results:

• The proposed detection algorithm demonstrated superior performance compared to existing methods on public datasets.
• The mask classification algorithm achieved an accuracy of 97.84% on the custom dataset, outperforming other algorithms.
• The fusion of transfer learning and deep learning enhanced model generalization and addressed overfitting issues.

Conclusions:

• The developed algorithm offers a robust solution for accurate, real-world face mask detection and classification.
• This technology can enhance public health measures by ensuring the use of effective personal protective equipment.
• The approach effectively tackles challenges posed by varied lighting, occlusion, and diverse mask types.