Scratch Train for Lightweight Models for Face Mask Detection

Automated systems for detecting face mask use in public became urgent during the COVID-19 pandemic. Most existing mask detection research fine-tunes ImageNet pre-trained backbones on relatively small datasets of masks. This approach raises concerns about model performance in situations with limited computational resources or when external pre-trained weights are not accessible. Additionally, there is a limited comparative analysis of recent lightweight architectures under consistent training conditions for mask detection tasks. This paper evaluates four state-of-the-art lightweight architectures for binary mask detection, including RepViT, ShuffleNetV2, EdgeNeXt-Small, and EfficientFormer. These models are trained from scratch using identical training protocols on two datasets containing 7,553 and 12,000 RGB images, respectively. The performances are then assessed using standardised metrics, including accuracy, precision, recall, and F1-score. Results demonstrate that ShuffleNet V2 achieves an optimal balance between classification accuracy and computational efficiency, delivering 0.987 accuracy on Dataset 1 and 0.998 accuracy on Dataset 2 while maintaining the fastest inference time of 0.464-0.667 milliseconds and the smallest model size of 1.26 million parameters. RepViT and EdgeNeXt-Small achieve slightly higher accuracy but require significant computational resources. EfficientFormer consistently underperforms across all evaluation metrics. These findings indicate that extremely lightweight CNNs can excel at mask detection when trained from scratch, making ShuffleNet V2 the ideal choice for resource-constrained deployment scenarios.