| Face Mask Detection for Real-Time Monitoring Using MobileNetV2: A Bias-Aware, Cross-Validated, and Efficient Deep Learning Framework |
| Paper ID : 1054-ICEEM2025 (R1) |
| Authors |
|
Wael Badawy *1, Ahmed Azouz2, Marwan Ragab2, Sherouk Ashraf2, Mariam Ashraf2, Manar Mahmoud2, Howaida Rabie2, Nadine Mohamed2 1Eru 2Egyptian Russian University |
| Abstract |
| This study presents a bias-aware, cross-validated, and efficiency-optimized face mask detection framework, integrating MobileNetV2 with transfer learning for real-time applications. The proposed system leverages an expanded dataset of 27,553 annotated images—comprising the original 7,553-image collection, a balanced 10,000-image subset from MaskedFace-Net, and 10,000 images from the Real-World Masked Face Dataset (RWMFD)—spanning diverse ethnicities, age groups, lighting conditions, and mask types. We implemented 5-fold cross-validation, external dataset testing, and demographic subgroup analysis (ethnicity, age, gender) to ensure robust generalization and fairness. An ablation study compares MobileNetV2 against MobileNetV3 and Efficient Net-Lite, while efficiency metrics (FLOPs, parameter count, latency, and memory usage) quantify trade-offs with accuracy. Results show that MobileNetV2 achieves an average F1-score of 98.7% across folds, 97.9% accuracy on unseen datasets, and consistent subgroup performance with less than 1.5% disparity across demographics. The model requires only 3.4M parameters and 299 MFLOPs, enabling real-time inference on low-power devices. This work demonstrates a deployable, equitable, and resource-efficient face mask detection system suitable for scalable public health monitoring. |
| Keywords |
| Face Mask Detection , MobileNetV2 , CNN , Real-Time Monitoring |
| Status: Accepted |