Abstract

Since the COVID-19 outbreak, face masks have played a crucial role in preventing virus transmission in public spaces. This project applies Faster R-CNN with a ResNet-50 + FPN backbone to develop an efficient facemask detection system. The model was trained on a custom dataset with three classes—mask, no mask, and false mask—using optimized hyperparameters (LR = 0.005, SGD). Evaluation on key metrics such as mAP, F1-score, and LAMR showed that the model achieved 98.40% mAP and nearly perfect detection for “false mask” class. The system was deployed with a PyQt GUI that supports both image and video inputs.

Key Contributions

• Created a customized 3-class facemask detection dataset in Pascal VOC format with 3954 labeled images.
• Designed and evaluated four Faster R-CNN backbones: MobileNetV2, VGG16, ResNet50, and ResNet50+FPN.
• Achieved highest performance using ResNet50+FPN (mAP = 98.40%) with optimized learning rate and SGD optimizer.
• Implemented interpretability via EigenCAM and AblationCAM to visualize model focus during prediction.
• Developed an intuitive GUI using PyQt with real-time mask detection in both image and video modes.

Conclusion

This project successfully demonstrates that Faster R-CNN with an enhanced backbone (ResNet50+FPN) is capable of accurately detecting mask-wearing status across various conditions. With robust performance and practical interpretability tools, the system offers real-world application potential in public health monitoring.

Limitation

The system's video detection operates at a low frame rate (≈1.5 fps), limiting real-time deployment efficiency. Additionally, performance on diverse camera angles and lighting conditions may require further generalization.

Future Work

Planned improvements include optimizing real-time video processing using lighter models (e.g., YOLOv7), deploying on edge devices, and extending the system for additional PPE detection (gloves, face shields) using multi-label object detection.