Deep Learning-Based Biomimetic Identification Method for Mask Wearing Standardization.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• The post-pandemic era necessitates efficient and accurate deep learning models for detecting mask-wearing compliance.
• Existing models lack standardization for detecting normalized mask-wearing, particularly concerning nose coverage.

Purpose of the Study:

• To enhance the accuracy and speed of deep learning models for detecting normalized mask-wearing.
• To address the research gap in standardized mask-wearing detection models, focusing on nose visibility.

Main Methods:

• Developed an improved YOLOv5s (You Only Look Once v5s) object detection model for mask-wearing normalization detection.
• Modified the BottleneckCSP module to BottleneckCSP-MASK and integrated an SE module for improved feature extraction of mask and nose targets.
• Optimized the bonding fusion layer for enhanced detection of mask and nose features.

Main Results:

• The proposed model achieved an overall detection accuracy of 99.3% with an average detection speed of 0.014 s/pic.
• Demonstrated effective detection of normalized and non-normalized mask-wearing across diverse individuals and complex backgrounds.
• The improved model showed a 0.5% increase in mAP compared to the original YOLOv5s and compressed model size by 10%.

Conclusions:

• The developed mask-wearing normalization detection model precisely identifies non-wearing, normalized wearing, and non-normalized wearing behaviors.
• The model's improvements in accuracy and speed make it suitable for real-world applications in the post-pandemic era.
• This research provides a standardized approach to mask-wearing detection, emphasizing nose visibility.