Fusion of LDA and PCA for Face Recognition Gian Luca Marcialis and Fabio Roli Department of Electrical and Electronic Engineering - University of Cagliari Piazza d’Armi - 09123 Cagliari (Italy) {marcialis,roli}@diee.unica.it Abstract. Although many approaches for face recognition have been proposed in the last years, none of them can overcome the main problem of this kind of biometrics: the huge variability of many environmental parameters (lighting, pose, scale). Hence, face recognition systems can achieve good results at the expense of robustness. In this work we describe a methodology for improving the robustness of a face recognition system based on the “fusion” of two well-known statistical representations of a face: PCA and LDA. Experimental results that confirm the benefits of fusing PCA and LDA are reported. 1 Introduction In the last years, Face Recognition [1] has become one of the most challenging task in the pattern recognition field. The recognition of faces is very important for many applications: video- surveillance, retrieval of an identity from a data base for criminal investigations and forensic applications. The face is considered a good biometric for many reasons: the acquisition process is non- intrusive and does not require collaboration of the subject to be recognised. The acquisition process of a face from a scene is simpler and cheaper than the acquisition of other biometrics as the iris and the fingerprint. On the other hand, many problems arise, because of the variability of many parameters: face expression, pose, scale, lighting, and other environmental parameters. For this reason, we can subdivide the applications which involve face recognition in two fields: applications in a controlled environment and applications in a uncontrolled environment. The first kind of applications refers to the problem of “identity authentication”: a subject submits to the system its face (frontal and/or profile view) and he declares his identity. The aim of the system is to verify the matching between the claimed identity and the given biometric. This kind of applications is typical for internet transactions, driver’s licenses, access to limited areas. The second kind of applications refers to the problem of “recognition of an identity in a scene”, and it is typical for video-surveillance applications. A system that automatically recognises a face in a scene, first detects it and normalises it with respect to the pose, lighting and scale. Then, the system tries to associate the face to one or more faces stored in its database, and gives the set of faces that are considered as “nearest” to the detected face. This problem is much more complex than the “verification” problem, and it requires more computational resources and very robust algorithms for detection, normalisation and recognition. Usually, each of these three problems is so complex that it must be studied separately. In this work, we propose some algorithms only for the face recognition problem. Our methodology is based on the fusion of multiple recognisers for improving the performance of the best individual one. The paper is structured as follows. In section 2 we briefly discuss the state of the art in face recognition systems and the role of multiple classifier systems. In section 3 we present our