Attention-enhanced U-net for robust N95 FFR segmentation in infrared imagery

ORCID

Clothilde Brochot : 0000-0002-2431-5053

Ali Bahloul : 0000-0002-4597-2001

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

Maison d’édition

IEEE

Résumé

Infrared (IR) imaging provides a non-invasive approach for evaluating the fit of N95 filtering facepiece respirators (N95 FFRs) and detecting potential leakage. This work presents an optimized U-Net architecture for N95 FFR segmentation from infrared images, a critical component for reliable leakage analysis. The proposed architecture integrates bilinear upsampling, Squeeze-and-Excitation (SE) blocks, and normalization techniques to improve feature representation, stabilize training, and reduce computational cost. The model is trained on a subset of participants and evaluated on separate validation and test partitions, composed of participants unseen during training and including mask types not encountered during that phase, enabling assessment of cross-participant generalization and robustness. Experimental results demonstrate that the proposed model achieves an IoU of 98.61% and a precision of 99.94%, outperforming the standard U-Net across all evaluated metrics while demonstrating strong generalization capability for practical healthcare and industrial applications.

Mots-clés

Équipement de protection respiratoire, Respirator, Masque N95, N95 mask

Numéro de projet IRSST

2022-0008

Ce document n'est pas disponible pour le moment.

Partager

COinS