Ammar Chouchane

Efficient Tensor-Based 2D+3D Face Verification

We propose a novel approach for face verification by encoding 2D and 3D face images as a high order tensor. To perform tensor dimensionality reduction for both the unsupervised and supervised cases, we propose multilinear whitened principal component analysis (MWPCA) and tensor exponential discriminant analysis (TEDA), respectively. MWPCA is utilized to solve the small sample size problem in the high-dimensional space and to improve the discrimination power achieved by classical MPCA. In the supervised case, we extend multilinear discriminant analysis to TEDA in order to emphasize the discriminant data included in the null space of the within-class scatter matrix of each tensor’s mode. Additionally, TEDA enlarges the margin between samples belonging to different classes via distance diffusion mappings. Our proposed approach can be seen as a novel data fusion method based on tensor representation. Indeed, the histograms of different local descriptors extracted from both 2D and 3D face modalities are combined through different tensor modes. The extensive experimental evaluation carried out on FRGC v2.0, Bosphorus, and CASIA 2D and 3D face databases indicates that the proposed approach performs significantly better than the state-of-the-art approaches.

3D face recognition in presence of expressions by fusion regions of interest

We propose a face recognition system insensitive to expressions. This system uses the fusion by concatenating the entire face with the regions of interest (nose, mouth, right eye and left eye). To enhance the discriminant information phases of Gabor filter are used. The Principal Component Analysis (PCA) + Enhanced Fisher linear discriminant Model (EFM) are applied to the data to find a reduced basis projection and discriminant. The classification is usually performed using a single distance measure in the final multidimensional space. In this work we use a support vector machine (SVM) architecture with one against all. The model is studied and applied to the CASIA color database and gives a recognition rate of overall evaluation RReval = 94.30% and the test set RRtest = 81.30%.

Local descriptors and tensor local preserving projection in face recognition

In this paper, a new multi-dimensional facial recognition system is proposed. A new technique for data reduction for multidimensional biometric facial analysis to improve face recognition performance in real environments is implemented. For this the tensorial methods are adopted, the sample of the face must be reshaped by natural tensor representations into vectors of very large dimensions. This remodeling breaks the natural structure of the correlations existing in the original tensor data, involving high costs and the need to evaluate a large number of parameters. Firstly, we give an overview and generalities on facial recognition systems, and then we present some techniques to n Dimensional Face Recognition System (nDFRS). The Tensor Local Preserving Projection (TLPP) is proposed as a new method of reducing and implemented to obtain our Nearest Neighbor classification. TLPP is used to reduce features vectors obtained by local descriptors LBP, LPQ and BSI. Many experiments on ORL, YALE and FERET Databases show that our methods are not only more effective but also more robust.

3D and 2D face recognition using integral projection curves based depth and intensity images

This paper presents an automatic face recognition system in the presence of illumination, expressions and pose variations based on depth and intensity information. At first, the registration of 3D faces is achieved using iterative closest point ICP. Nose tip point must be located using Maximum Intensity Method. This point usually has the largest depth value; however there is a problem with some unnecessary data such as: shoulders, hair, neck and parts of clothes; to cope with this issue, we propose the integral projection curves IPC-based facial area segmentation to extract the facial area. After that, the combined method principal component analysis PCA with enhanced Fisher model EFM is used to obtain the feature matrix vectors. Finally, the classification is performed using distance measurement and support vector machine SVM. The experiments are implemented on two face databases CASIA3D and GavabDB; our results show that the proposed method achieves a high recognition performance.

3D face recognition based on histograms of local descriptors

Face recognition in an uncontrolled condition such as illumination and expression variations is a challenging task. Local descriptor is one of the most efficient methods used to deal with these problems. In this paper, we present an automatic 3D face recognition approach based on three local descriptors, local phase quantization (LPQ), Three-Patch Local Binary Patterns (TPLBP) and Four-Patch Local Binary Patterns (TPLBP). Facial images are passing through one of the three descriptors and divided into sub-regions or rectangular blocks. The histogram of each sub-region is extracted and concatenated into a single feature vector. PCA (Principal Component Analysis) and EFM (Enhanced Fisher linear discriminant Model) are used to reduce the dimensionality of the resulting feature vectors. Finally, these vectors are sent to the classification step, when we use two methods; SVM (Support Victor Machine) and similarity measures. CASIA 3D face database is introduced to experimental evaluation. The experimental results illustrate a high recognition performance of the proposed approach.

A Novel Hybrid Approach for 3D Face Recognition Based on Higher Order Tensor

This paper presents a new hybrid approach for 3D face verification based on tensor representation in the presence of illuminations, expressions and occlusion variations. Depth face images are divided into sub-region and the Multi-Scale Local Binarised Statistical Image Features (MSBSIF) histogram are extracted from each sub-region and arranged as a third order tensor. Furthermore, to reduce the dimensionality of this tensor data, we use a novel hybrid approach based on two steps of dimensionality reduction multilinear and non-linear. Firstly, Multilinear Principal Component Analysis (MPCA) is used. MPCA projects the input tensor in a new lower subspace in which the dimension of each tensor mode is reduced. After that, the non-linear Exponential Discriminant Analysis (EDA) is used to discriminate the faces of different persons. Finally, the matching is performed using distance measurement. The proposed approach (MPCA+EDA) has been tested on the challenging face database Bosporus 3D and the experimental results show that our method achieves a high verification performance compared with the state of the art.

Multilinear Side-Information based Discriminant Analysis for Face and Kinship Verification in the Wild

Abstract This paper presents a new approach for face and kinship verification under unconstrained environments. The proposed approach is based on high order tensor representation of face images. The face tensor is built based on local descriptors extracted at multiscales. Besides, we formulate a novel Multilinear Side-Information based Discriminant Analysis (MSIDA) to handle the weakly supervised multilinear subspace projection and classification. Using only the weak label information, MSIDA projects the input face tensor in a new subspace in which the discrimination is improved and the dimension of each tensor mode is reduced simultaneously. Experimental evaluation on four challenging face databases (LFW, Cornell KinFace, UB KinFace and TSKinface) demonstrates that the proposed approach significantly outperforms the current state of the art.

3D face verification across pose based on euler rotation and tensors

In this paper, we propose a new approach for 3D face verification based on tensor representation. Face challenges, such as illumination, expression and pose, are modeled as a multilinear algebra problem where facial images are represented as high order tensors. Particularly, to account for head pose variations, several pose scans are generated from a single depth image using Euler transformation. Multi-bloc local phase quantization (MB-LPQ) histogram features are extracted from depth face images and arranged as a third order tensor. The dimensionality of the tensor is reduced based on the higher-order singular value decomposition (HOSVD). HOSVD projects the input tensor in a new subspace in which the dimension of each tensor mode is reduced. To discriminate faces of different persons, we utilize the Enhanced Fisher Model (EFM). Experimental evaluations on CASIA-3D database, which contains large head pose variations, demonstrate the effectiveness of the proposed approach. A verification rate of 98.60% is obtained.

2D and 3D face recognition based on IPC detection and patch of interest regions

In this paper, we propose a framework of Face Recognition System (FRS). Essentially, we are focused on the face detection process and the role of interest regions of the human face. In order to locate exactly the facial area, we propose the use of horizontal and vertical IPC (Integral Projection Curves). The role of important patches of face: nose and eyes is investigated in this work. An efficient method based on PCA (Principal component analysis) followed by EFM (Enhanced Fisher Model) is used to build the characteristic features, these latter are sent to the classification step using two methods, Distance Measurements and SVM (Support Vector Machine). Finally, the effect of fusion of two modalities (2D and 3D) is studied and examined. Experiments are performed on the CASIA3D face database which contains 123 persons under varying of illumination, expression variation.

Multimodal face recognition based on histograms of three local descriptors using score level fusion

In this paper, we propose an efficient framework of multimodal face recognition that explores 2D and 3D information based on the score level fusion. To solve the problems of illumination and expression variations, three local methods are introduced, Local Phase Quantization (LPQ), Three-Patch Local Binary Patterns (TPLBP) and Four-Patch Local Binary Patterns (TPLBP). After applying local descriptors to the input image (2D and 3D), this latter is divided into sub-regions or rectangular blocks. Then, the histogram of each sub-region is extracted and concatenated into a single features vector. Principal Component Analysis (PCA) and Enhanced Fisher linear discriminate Model (EFM) are used to reduce the dimensionality. Classification is then performed using the robust Support Vector Machine (SVM) classifier. Finally, score level fusion is used to improve the recognition performance. Experiments are implemented on CASIA3D face database. Our results show that the proposed approach achieves very high performance with RR=98.65% and EER=0.67%.