| Titre : | Facial Recognition in Low-Quality Images |
| Auteurs : | Sana BELLILI, Auteur ; Abdelmalik Ouamane, Auteur |
| Type de document : | Monographie imprimée |
| Editeur : | Université Mohamed Kheider, 2026 |
| Langues: | Anglais |
| Mots-clés: | Face recognition, Low-resolution face verification, Super-resolution, Multilinear subspace learning, Tensor Cross-View Quadratic Discrimi nant Anal?ysis, Vision Transformers, feature enhancement, Optimized model SR-Vit, Met?ric learning, Biometric recognition. |
| Résumé : |
Low-resolution face verification remains a major challenge in computer vision and
biometrics, particularly in real-world surveillance scenarios where images suffer from poor resolution, noise, blur, and illumination variations. These degrada?tions significantly reduce identity-discriminative features and limit the robust?ness of conventional recognition systems. This work proposes a comprehensive framework that combines multilinear subspace learning and Vision Transformer models enhanced by super-resolution techniques to improve verification perfor?mance under extreme low-resolution conditions. First, a tensor-based verification framework is introduced using multilinear repre?sentations and discriminant tensor learning. Advanced metric learning methods including TXQDA, MSIDA, SILD and XQDA are integrated within a multi?scenario verification scheme, where TXQDA and MSIDA belong to multilinear subspace learning, while XQDA and SILD are linear subspace learning methods used for comparison and complementary evaluation. TXQDA preserves the in?trinsic multi-dimensional structure of facial data without vectorization, leading to stronger discriminative power and improved robustness. Experiments are con?ducted on extremely low-resolution scales (16×16, 32×32, and 48×48). Super?resolution enhancement using SRResNet, SRGAN, and Real-ESRGAN with a ×4 upscaling factor further boosts recognition accuracy. The proposed approach achieves verification rates of 99.17% on LFW, 90.90% on CelebA, and 76.90% on QMUL-SurvFace. Second, the study investigates Vits for low-resolution face verification by evalu?ating multiple architectures including SimpleViT, DistillableViT, DeepViT, and CaiT combined with super-resolution preprocessing using SRResNet and SR?GAN. Evaluations on QMUL-TinyFace and QMUL-SurvFace show that super?resolution significantly improves transformer-based verification, with Simple?ViT and DistillableViT achieving near-perfect accuracy. The proposed SR-ViT framework surpasses state-of-the-art methods, reaching 52.92% TAR@FAR=1% and 92.65% AUC on QMUL-TinyFace, with strong AUC gains on QMUL?SurvFace. |
| Type de document : | Thése doctorat |
Disponibilité (1)
| Cote | Support | Localisation | Statut | Emplacement | |
|---|---|---|---|---|---|
| Th/1451 | Thèse de doctorat | BIB.FAC.ST. | Empruntable | Magazin |
Erreur sur le template
