AdaBoost Multiple Feature Selection and Combination for Face Recognition Francisco Mart´ ınez-Contreras, Carlos Orrite-Uru˜ nuela, and Jes´ us Mart´ ınez-del-Rinc´on CVLab, Aragon Institute for Engineering Research, University of Zaragoza, Spain {franjmar,corrite,jesmar}@unizar.es Abstract. Gabor features have been recognized as one of the most suc- cessful face representations. Encouraged by the results given by this ap- proach, other kind of facial representations based on Steerable Gaussian first order kernels and Harris corner detector are proposed in this pa- per. In order to reduce the high dimensional feature space, PCA and LDA techniques are employed. Once the features have been extracted, AdaBoost learning algorithm is used to select and combine the most representative features. The experimental results on XM2VTS database show an encouraging recognition rate, showing an important improve- ment with respect to face descriptors only based on Gabor filters. 1 Introduction Face recognition has become in a relevant biometrics modality due to the broad variety of potential applications in public security, law enforcement, access con- trol and video surveillance. The main two factors that affect to the performance of most current systems are the large variability in facial appearance of a person, due to different poses, lighting and facial expressions, and the high dimension- ality of the facial features. In order to deal with facial expression and illumination, some authors [8,9] propose to use a Gabor wavelet representation. The Gabor wavelet kernels al- low to capture salient visual properties such as spatial localization, orientation selectivity, and spacial frequency characteristics. Recently, others authors [21,22] proposed the use of local oriented Gaussian derivative filters for face identification. The structure of an image can be repre- sented by the outputs of a set of multi-scales Gaussian derivative filters (MGDF) applied to the image as proposed in [15], for the appearance representation in face recognition and image retrieval. More biologically motivated is the use of Steerable Gaussian first order derivative filters as proposed in [4] for detecting facial landmarks from neutral and expressive facial images. The main drawback when using Gabor or other oriented filter approaches to obtain a new representation of the face is the great redundancy and high dimen- sionality they exhibit. Recent papers [19,17] have proposed using the AdaBoost learning algorithm presented in [18] to select a small set of filters. H. Araujo et al. (Eds.): IbPRIA 2009, LNCS 5524, pp. 338–345, 2009. c  Springer-Verlag Berlin Heidelberg 2009