Transfer learning based cascaded deep learning network and mask recognition for COVID-19.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• The ongoing COVID-19 pandemic necessitates effective public health measures, including mask-wearing enforcement in public spaces.
• Existing mask detection systems face challenges with non-standard mask types (e.g., scarves, cotton masks), leading to potential bypasses.
• Accurate identification of mask presence and type is crucial for public health surveillance and safety.

Purpose of the Study:

• To develop an advanced mask recognition system capable of detecting both the presence and type of masks worn by pedestrians.
• To enhance the accuracy and robustness of mask detection systems by employing a cascaded deep learning approach.
• To leverage transfer learning and lightweight network architectures for efficient and effective mask recognition.

Main Methods:

• Proposed a cascaded deep learning network architecture based on the MobilenetV3 lightweight network.
• Modified MobilenetV3 networks by altering the output layer activation function and model structure for cascading.
• Integrated transfer learning with a multi-task convolutional neural network (MTCNN) for face detection and two modified MobilenetV3 networks for mask feature extraction.
• Trained the cascaded network using pre-trained ImageNet parameters to reduce computational load.

Main Results:

• The cascaded deep learning network achieved a 7% improvement in classification accuracy compared to individual modified MobilenetV3 networks.
• The system demonstrated excellent performance in recognizing mask presence and types.
• The integration of MTCNN for face detection and cascaded MobilenetV3 for mask feature extraction proved effective.

Conclusions:

• The proposed cascaded deep learning network offers a significant improvement in mask recognition accuracy.
• This system provides a robust solution for detecting various types of masks, addressing limitations of previous systems.
• The developed mask recognition system can be effectively deployed in public spaces to enhance safety and compliance with health guidelines.