Robust patch-level infrared leak detection on N95 respirators using spatiotemporal deep learning
Type de document
Articles dans des actes de congrès
Année
2026
Langue
Anglais
Directeurs de la publication
The Institute of Electrical and Electronics Engineers
Titre des actes
2026 18th International Conference on Electronics, Computers and Artificial Intelligence (ECAI)
Maison d’édition
IEEE
Résumé
Face seal leaks are a major cause of protection loss for N95 filtering facepiece respirators (N95 FFR), yet current fit-testing methods provide only a global fit factor and no information about leak location. Infrared (IR) imaging offers a way to visualise leaks by capturing temperature variations near the mask seal, but most existing approaches operate at the level of full-face images and do not explicitly model the temporal dynamics of breathing or the strong class imbalance between leak and non-leak samples. In this work, controlled infrared videos of N95 FFRs under calibrated leak scenarios, acquired on a respiratory test bench, are used to construct patch-level sequences centred on leak locations along the seal. A Bi-Level Class Balancing (BLCB) strategy is introduced, combining global resampling with targeted preservation of hard negative sequences. Three spatiotemporal deep architectures are evaluated, including two CNN-LSTM models and an attention-based model, all trained with a focal-type loss on balanced sequence sets. Experiments show that the attention-based model achieves high sequence-level performance (up to 98% accuracy and AUC-ROC close to 0.99 on several regions), while the BLCB scheme effectively mitigates class imbalance. Cross-device evaluation further indicates that models trained on one configuration maintain strong discriminative ability when transferred to the other, with only a modest drop in accuracy and F1-score. These results demonstrate the potential of patch-level spatiotemporal modelling combined with dedicated class balancing for accurate and robust infrared leak detection on N95 FFRs.
Mots-clés
Équipement de protection respiratoire, Respirator, Masque N95, N95 mask
Numéro de projet IRSST
2022-0008
Citation recommandée
Baccari, R., Benmoussa, H., Yaddaden, Y., Arbane, M., Brousseau, J., Brochot, C., . . . Maldague, X. (2026). Robust patch-level infrared leak detection on N95 respirators using spatiotemporal deep learning. Dans The Institute of Electrical and Electronics Engineers (édit.), 2026 18th International Conference on Electronics, Computers and Artificial Intelligence (ECAI). IEEE. https://doi.org/10.1109/ECAI69016.2026.11613597