Face Authentication Using Enhanced Fisher linear discriminant Model (EFM) DJAMEL SAIGAA 1 , N. BENOUDJIT 2 , K. BENMAHAMED 3 , S LELANDAIS 4 1 Automatics Department, University Mohamed khider B.P 145 RP Biskra (07000), ALGERIA, 2 Electronics Department, University of Batna (05000), ALGERIA 3 Electronics Department, University of Setif (19000), ALGERIA, 4 Complex Systems Laboratory (LSC), University of Evry Val Essonne, FRENCH Abstract: - In this paper, Enhanced Fisher linear discriminant Model (EFM) is presented as an alternative feature extraction algorithm to Principal Component Analysis (PCA) widely used in automatic face recognition/authentication tasks. We show that the promising EFM algorithm extracts from faces features that are relevant and efficient for authentication. This leads to improved success rates and a reduced client model size over a PCA based feature extraction. The feasibility of the EFM method has been successfully tested on face authentication using 2360 XM2VTS frontal face images corresponding to 295 subjects, which were acquired under variable illumination and facial expressions. By the EFM method we obtain an equal error rate of 1.96% on face authentication using only 56 features. Key-Words: - Eigenfaces, Enhanced Fisher linear discriminant Model (EFM), Face authentication, Fisher Linear Discriminant (FLD), Principal Component Analysis (PCA). 1 Introduction Automatic personal identity verification based on facial images is important for many security applications [1] [2] [3]. In face authentication, as in most image processing problems, features are extracted from the images before processing. Working with rough images is not efficient: in face authentication, several images of a single person may be dramatically different, because of changes in viewpoint, in colour and illumination, or simply because the person's face looks different from day to day. Therefore extracting relevant features, or discriminant ones, is a must. Nevertheless, one hardly knows in advance which possible features will be discriminant or not. For this reason, one of the methods often used to extract features in face authentication is PCA (Principal Component Analysis) [4]. Another family of methods is the local features based methods such as [5], or those based on FLD (Fisher Linear Discriminant) as in [6]. In this paper, we show how the promising EFM (Enhanced Fisher linear discriminant Model) technique extracts features that are more closely related to our intuition of discriminant information, and that improve the success rate compared to an equivalent system using PCA or FLD. EFM also belongs to the family of subspace methods [7]. The remaining of this paper is organised as follows. Section 2 presents the problem of face authentication. Section 3 shows how to extract features from rough images, and presents the procedure based on EFM. Section 4 shows the experimental results. Section 5 concludes the paper. 2 Face authentication Face authentication systems typically compare a feature vector X extracted from the face image to verify with a client template, consisting in similar feature vectors i Y extracted from images of the claimed person stored in a database ( 1 i p ≤ ≤ , where p is the number of images of this person in the learning set). The matching may be made in different ways, one being to take the Euclidean distance between vectors (this method will be taken as an example here). If the distance between X and i Y is lower than a threshold, the face from which X is extracted will be deemed to correspond 4th WSEAS Int. Conf. on COMPUTATIONAL INTELLIGENCE, MAN-MACHINE SYSTEMS and CYBERNETICS Miami, Florida, USA, November 17-19, 2005 (pp155-160)