Skip to Main content Skip to Navigation

‫Reconnaissance Biométrique par Fusion Multimodale du Visage 2D et 3D

Abstract : Facial recognition is one of the best biometric modalities for applications related to the identification or verification of people. Indeed, it is the modality used by humans. It is non-intrusive and socially well accepted. Unfortunately, human faces are similar and therefore offer little possibility of distinction from other biometric modalities, such as fingerprints and iris. Moreover, when it comes to 2D images of faces, the intra-class variations, due to factors as diverse as the changes in lighting conditions, variation of cosmetics and pose, are generally greater than the inter-class variations. classes, which makes 2D face recognition unreliable under real-world conditions. Recently, 3D representations of faces have been widely studied by the scientific community to address unresolved issues in 2D facial recognition. This thesis is devoted to robust facial recognition using 2D and 3D facial data fusion. We devote the first part of our study to uni-modal face verification and 2D multi-face algorithms. First, we study several methods to select the best face authentication systems. Next, we present multi-modality and score fusion methods for both combination and classification approaches. Finally, merging methods of scores are compared on the XM2VTS face database. In the second part, we propose an automatic face authentication algorithm by merging two multimodal systems (multi-algorithms and multi-sensors 2D + 3D). First, we corrected the rotation of the head by the ICP algorithm and then presented six local feature extraction methods (MSLBP, proposed CSL, Gabor Wavelets, LBP, LPQ and BSIF). The classification of the characteristics is carried out by the cosine metric after reduction of space by EFM, then fusion at scores level by a good classifier with two classes SVM. Finally, the application is performed on the CASIA 3D and Bosphorus databases. In the last part, we study the uni-modal 2D and 3D and multimodal (2D + 3D) face verification based on the fusion of local information. The study consists of three main stages (preprocessing, feature extraction and classification). In the first step, a preprocessing phase is necessary. The ICP algorithm is used to align all faces and the PS approach is used to reduce the influence of dimming for 2D images. In the second step, we used four local descriptors (LBP, LPQ, BSIF and proposed Statistical LBP). After extracting the features, the 2D or 3D facial image is divided into 10 regions and each region is divided into 15 small blocks. The extraction of local characteristics is summarized by the corresponding histograms of each block. In the last step, we propose to use EDA coupled to WCCN for histogram dimension reduction for each region. We validate our proposed methods by comparing them with those existing in the scientific literature on the FRGC v2 and CASIA 3D databases.
Complete list of metadatas

Cited literature [389 references]  Display  Hide  Download
Contributor : Abdelmalik Ouamane <>
Submitted on : Wednesday, August 1, 2018 - 4:10:10 PM
Last modification on : Monday, August 6, 2018 - 12:03:14 PM
Long-term archiving on: : Friday, November 2, 2018 - 2:23:39 PM


ouamane abdelmalik.pdf
Files produced by the author(s)


  • HAL Id : tel-01852447, version 1


Abdelmalik Ouamane. ‫Reconnaissance Biométrique par Fusion Multimodale du Visage 2D et 3D. Vision par ordinateur et reconnaissance de formes [cs.CV]. University Mohamed Khider Biskra, Algeria, 2015. Français. ⟨tel-01852447⟩



Record views


Files downloads