A multi-class framework for face mask compliance detection using lightweight deep learning models

Balraj E; Manikandan P; Sambath M; Omana J

doi:10.1038/s41598-026-50603-1

A multi-class framework for face mask compliance detection using lightweight deep learning models

Sci Rep. 2026 May 19. doi: 10.1038/s41598-026-50603-1. Online ahead of print.

Authors

Balraj E¹, Manikandan P², Sambath M², Omana J²

Affiliations

¹ School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India. balraj.e@vit.ac.in.
² School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India.

PMID: 42156444
DOI: 10.1038/s41598-026-50603-1

Abstract

The use of face mask compliance regarding compliance monitoring is important in health sensitive and surveillance driven environments. This paper has introduced a multi-class face mask compliance detection system in the form of a Deep Convolutional Neural Network (DCNN) and advanced MobileNetV3 model. The proposed system is also based on a more difficult three-class problem (correct mask use, improper mask use, and no mask) as opposed to the traditional binary mask detection methods. A balanced custom curated dataset from KAGGLE of 2,079 images was experimented on, manually checked to be of quality, consistent, and equally represent all the classes, including improper mask-wearing cases that are usually underrepresented in prior studies.In the three-class task, the standard MobileNetV3 architecture was recast by substituting the final classification head with a lightweight fully connected layer of 256 neurons paired with a three-neuron softmax output layer. Squeeze-and-Excitation (SE) blocks were added to make the channel-wise recalibration of features and better separation of visually related classes. Stratified hold-out validation strategy (a 70:15:15 train-validation-test split) was employed to guarantee that performance is assessed by reliably generalizing. we were able to perform performance evaluation through accuracy, precision, recall, F1-score, specificity, Cohen Kappa, and confusion matrix analysis. Besides the comparison of DCNN, the given model was compared with the representative lightweight and state-of-the-art frameworks that include the MobileNetV2-, VGG16-, and YOLO-based models. MobileNetV3 obtained 98.90% accuracy and F1-score of 0.989, which proves to be competitive or even better in a more challenging multi-class compliance environment. In the future, the emphasis will be put on cross-dataset validation, real-time video monitoring, and the integration of explainable AI in order to improve practical implementation.

Keywords: COVID-19; DCNN; Deep Learning; MobileNet V3; Multi-Class Mask Detection; Public Health Monitoring; Squeeze and Excitation Block.