Identification Biométrique par Reconnaissance Faciale en Lumière Visible et Infrarouge / BENAMARA Nadir Kamel 

Public ISBD Titre : Identification Biométrique par Reconnaissance Faciale en Lumière Visible et Infrarouge Type de document : document électronique Auteurs : BENAMARA Nadir Kamel, Auteur Année de publication : 2021-2022 Accompagnement : CD Langues : Français (fre) Catégories : Génie électrique:Vision et Technologies de l’Information et de la Communication Mots-clés : Reconnaissance Faciale, Imagerie Infrarouge, Sélection de Caractéristiques Aléatoire (RFS), Réseau Adverse Génératif Cyclique Thermique-Visible (TV-CycleGAN). : Face Recognition, Infrared Imagery, Random Feature Selection (RFS), Thermal-Visible Cycle Generative Adversarial Network (TV-CycleGAN). Résumé : La reconnaissance faciale correspond au procédé d’identification le plus naturel dans la communication humaine. Parmi les avantages offerts par la technologie biométrique basée sur la modalité faciale, sa performance assurée dans des environnements contrôlés, non intrusive pour une utilisation dans des endroits très peuplés et son coût accessible grâce aux récents développements massifs des capteurs CMOS/CCD. Les systèmes de reconnaissance faciale classiques reposent sur des images collectées à partir de capteurs utilisant le spectre visible, allant de 0.4 µm à 0.8 µm. Des limitations sont observées pour ces systèmes, principalement le changement de l’intensité et la qualité de l’éclairage qui détériorent leur performance de reconnaissance, voire inexploitables pour une utilisation nocturne. D’autres problèmes sont constatés pour cette composante spectrale comme les expressions faciales, les variations dans la pose ainsi que les déguisements et l’usurpation. Etant invariant aux changements d’éclairage, le spectre infrarouge apporte des atouts considérables pour améliorer les performances des systèmes de reconnaissance faciale. Cependant, l’imagerie infrarouge dispose aussi d’autres limitations comme l’opacité des lunettes et la perte de l’information texturale du visage dans sa composante thermique, contrairement au spectre visible. Face recognition is the most natural identification process in human communication. The advantages offered by facial biometry technology can be summarized in a guaranteed performance in controlled environments, suitable for use in highly populated places since it is non-intrusive and costly effective thanks to the recent massive developments of CMOS/CCD sensors. Conventional facial recognition systems rely on images collected by sensors that use the visible spectrum, ranging from 0.4µm to 0.8µm. These systems suffer from several limitations, mainly the intensity and the quality changes of lighting that decrease their recognition performance, even unusable for night applications. Other problems are considered for this spectral range such as facial expressions, pose variations as well as disguises and spoofing. Being invariant to light changes, the infrared spectrum brings a considerable asset to improve facial recognition performances. However, infrared imaging also has other challenges, such as the opacity of glasses and the loss of the face’s textures in its thermal sub-band, unlike the visible spectrum. In this thesis, we propose to take advantage of the two spectra, the visible and the infrared, to improve current facial recognition systems. Two approaches are considered for this purpose, multispectral recognition and heterogeneous recognition. The first approach aims to gather discriminative properties through a fusion scheme based on the simultaneous use of the visible and infrared spectra. A multispectral face recognition system has been developed for this purpose by combining the extracted feature vectors from each spectral sub-band, whereas a proposed algorithm developed for the best feature selection optimizes the resulting vector. In addition, the classification stage of this system is optimized by the particle swarm optimization metaheuristic to find the best parameters for the classifier. The second approach meets the need for recognition when the acquisition system is based on a different spectral modality than the one used to collect the gallery samples during the enrolment stage. In this context, a recognition system has been proposed based on a cycle generative adversarial network named TV-CycleGAN to synthesize facial images in the visible spectrum from the ones collected by thermal sensors. Finally, synthesized images are matched with those of the gallery for recognition. The experimental results obtained by our proposed recognition systems using visible light and infrared demonstrate their efficacy against those reported in the literature. Directeur de thèse : BOUDGHENE STAMBOULI Tarik Identification Biométrique par Reconnaissance Faciale en Lumière Visible et Infrarouge [document électronique] / BENAMARA Nadir Kamel, Auteur . - 2021-2022 . -  + CD. Langues : Français (fre) Catégories : Génie électrique:Vision et Technologies de l’Information et de la Communication Mots-clés : Reconnaissance Faciale, Imagerie Infrarouge, Sélection de Caractéristiques Aléatoire (RFS), Réseau Adverse Génératif Cyclique Thermique-Visible (TV-CycleGAN). : Face Recognition, Infrared Imagery, Random Feature Selection (RFS), Thermal-Visible Cycle Generative Adversarial Network (TV-CycleGAN). Résumé : La reconnaissance faciale correspond au procédé d’identification le plus naturel dans la communication humaine. Parmi les avantages offerts par la technologie biométrique basée sur la modalité faciale, sa performance assurée dans des environnements contrôlés, non intrusive pour une utilisation dans des endroits très peuplés et son coût accessible grâce aux récents développements massifs des capteurs CMOS/CCD. Les systèmes de reconnaissance faciale classiques reposent sur des images collectées à partir de capteurs utilisant le spectre visible, allant de 0.4 µm à 0.8 µm. Des limitations sont observées pour ces systèmes, principalement le changement de l’intensité et la qualité de l’éclairage qui détériorent leur performance de reconnaissance, voire inexploitables pour une utilisation nocturne. D’autres problèmes sont constatés pour cette composante spectrale comme les expressions faciales, les variations dans la pose ainsi que les déguisements et l’usurpation. Etant invariant aux changements d’éclairage, le spectre infrarouge apporte des atouts considérables pour améliorer les performances des systèmes de reconnaissance faciale. Cependant, l’imagerie infrarouge dispose aussi d’autres limitations comme l’opacité des lunettes et la perte de l’information texturale du visage dans sa composante thermique, contrairement au spectre visible. Face recognition is the most natural identification process in human communication. The advantages offered by facial biometry technology can be summarized in a guaranteed performance in controlled environments, suitable for use in highly populated places since it is non-intrusive and costly effective thanks to the recent massive developments of CMOS/CCD sensors. Conventional facial recognition systems rely on images collected by sensors that use the visible spectrum, ranging from 0.4µm to 0.8µm. These systems suffer from several limitations, mainly the intensity and the quality changes of lighting that decrease their recognition performance, even unusable for night applications. Other problems are considered for this spectral range such as facial expressions, pose variations as well as disguises and spoofing. Being invariant to light changes, the infrared spectrum brings a considerable asset to improve facial recognition performances. However, infrared imaging also has other challenges, such as the opacity of glasses and the loss of the face’s textures in its thermal sub-band, unlike the visible spectrum. In this thesis, we propose to take advantage of the two spectra, the visible and the infrared, to improve current facial recognition systems. Two approaches are considered for this purpose, multispectral recognition and heterogeneous recognition. The first approach aims to gather discriminative properties through a fusion scheme based on the simultaneous use of the visible and infrared spectra. A multispectral face recognition system has been developed for this purpose by combining the extracted feature vectors from each spectral sub-band, whereas a proposed algorithm developed for the best feature selection optimizes the resulting vector. In addition, the classification stage of this system is optimized by the particle swarm optimization metaheuristic to find the best parameters for the classifier. The second approach meets the need for recognition when the acquisition system is based on a different spectral modality than the one used to collect the gallery samples during the enrolment stage. In this context, a recognition system has been proposed based on a cycle generative adversarial network named TV-CycleGAN to synthesize facial images in the visible spectrum from the ones collected by thermal sensors. Finally, synthesized images are matched with those of the gallery for recognition. The experimental results obtained by our proposed recognition systems using visible light and infrared demonstrate their efficacy against those reported in the literature. Directeur de thèse : BOUDGHENE STAMBOULI Tarik

Exemplaires

Code-barres	Cote	Support	Localisation	Section	Disponibilité
1959	02-09-624	Version numérique et papier	Bibliothèque USTOMB	Thèse de Doctorat	Exclu du prêt

Documents numériques

02-09-624.pdf Adobe Acrobat PDF