Reconnaissance Biométrique par Fusion Multimodale de Visages

ADEL, SAOUD (2022) Reconnaissance Biométrique par Fusion Multimodale de Visages. Doctoral thesis, Faculté des Sciences et de la technologie.

Text thèse2021Adel saoud.pdf Download (3MB)

Abstract

Biometric systems are considered to be one of the most effective methods of protecting and securing private or public life against all types of theft. Facial recognition is one of the most widely used methods, not because it is the most efficient and reliable, but rather because it is natural and non-intrusive and relatively accepted compared to other biometrics such as fingerprint and iris. The goal of developing biometric applications, such as facial recognition, has recently become important in smart cities. Over the past decades, many techniques, the applications of which include videoconferencing systems, facial reconstruction, security, etc. proposed to recognize a face in a 2D or 3D image. Generally, the change in lighting, variations in pose and facial expressions make 2D facial recognition less than reliable. However, 3D models may be able to overcome these constraints, except that most 3D facial recognition methods still treat the human face as a rigid object. This means that these methods are not able to handle facial expressions. In this thesis, we propose a new approach for automatic face verification by encoding the local information of 2D and 3D facial images as a high order tensor. First, the histograms of two local multiscale descriptors (LPQ and BSIF) are used to characterize both 2D and 3D facial images. Next, a tensor-based facial representation is designed to combine all the features extracted from 2D and 3D faces. Moreover, to improve the discrimination of the proposed tensor face representation, we used two multilinear subspace methods (MWPCA and MDA combined with WCCN). In addition, the WCCN technique is applied to face tensors to reduce the effect of intra-class directions using a normalization transform, as well as to improve the discriminating power of MDA. Our experiments were carried out on the three largest databases: FRGC v2.0, Bosphorus and CASIA 3D under different facial expressions, variations in pose and occlusions. The experimental results have shown the superiority of the proposed approach in terms of verification rate compared to the recent state-of-the-art method.

Actions (login required)

View Item