COVID-19 face mask usage has sparked interest in improving the speed and accuracy of detecting masked faces in intricate surroundings. One approach is using deep learning methods like the YoloX network, but it suffers from the loss of semantic information and confounding effects that affect its precise localization capability. Although the Channel Enhancement Feature Pyramid (CE-FPN) can better mitigate the confounding effect, its own structure causes more computation and number of parameters. In order to address these issues, we propose an improved channel-enhanced feature pyramid (ICE-FPN) that is able to reduce the computational and parametric counts of CE-FPN and mitigate the confounding effect. Then, feature enhancement modules (FEM and FEM-s) are introduced at the neck of the network to enhance the semantic richness of feature fusion in the neck network, while contextual attention enhancement modules (CAEM) are added at the end of the neck to effectively improve the precise localization capability for mask-wearing targets in complex contexts. Finally, we utilize the technique of transfer learning during our training process, and the ablation experiments performed on the WMD dataset and the comparison experiments performed on the PWMFD dataset show a significant performance improvement of our proposed ICE-YoloX, and the results show that mAP0.5 improves from 99.54 to 99.62% and mAP0.75 improves from 89.47% in the WMD dataset improved to 91.64%.
- Deep Learning
- Face mask detection