MutualBoost learning for selecting Gabor features for face recognition Linlin Shen, Li Bai * School of Computer Science and IT, University of Nottingham, Nottingham NG8 1BB, UK Available online 17 April 2006 Abstract This paper describes an improved boosting algorithm, the MutualBoost algorithm, and its application in developing a fast and robust Gabor feature based face recognition system. The algorithm uses mutual information to eliminate redundancy among Gabor features selected using the AdaBoost algorithm. Selected Gabor features are then subjected to Generalized Discriminant Analysis (GDA) for class separability enhancement before being used for face recognition. Compared with one of the top performers in the 2004 face verification competition, our method demonstrates clear advantages in classification accuracy, memory and computation. The method has been tested on the whole FERET database using the FERET evaluation protocol. Significant improvement in performance is observed. For example, existing Gabor based methods use a huge number of Gabor features, our method needs only hundreds of Gabor features to achieve very high classification accuracy. Due to substantially reduced feature dimension, memory and computation costs are reduced significantly – only 4 s are needed to recognize 200 face images. Ó 2006 Elsevier B.V. All rights reserved. Keywords: Gabor filters; AdaBoost algorithm; Generalized discriminant analysis 1. Introduction Lades et al. (1993) pioneered the application of Gabor filters for face recognition. They proposed the Dynamic Link Architecture (DLA), which recognizes faces by extracting Gabor jets at each node of a rectangular grid over the face image. Wiskott et al. (1997) extended DLA and proposed the Elastic Bunch Graph Matching (EBGM) algorithm. This again stores local facial features in a graph data structure. The EBGM algorithm was the top per- former in the FERET evaluation contest. However, both DLA and EBGM require extensive amounts of computa- tion (Wiskott et al., 1997) – 30 s on a SPARC station 10- 512. Gabor filters can also be used to extract features from the face image as a whole. For example, in (Liu and Wechs- ler, 2002) a face image is first convolved with 40 Gabor fil- ters, the extracted high dimensional (160,000+) Gabor features are then subjected to linear discriminant analysis for dimension reduction. Similar work can be found in (Ayinde and Yang, 2002). Shen and Bai (2004b) extended the Gabor feature space to the kernel space to achieve better classification performance than linear dimension reduction methods. Their method was one of the top two performers in the 2004 face verification competition (Messer et al., 2004). Despite robustness, Gabor filter based feature selection methods are normally computationally expensive due to high dimensional Gabor features. To reduce feature dimen- sion, a sampling method is proposed in (Liu et al., 2004) to select ‘optimal’ positions on a face to extract Gabor fea- tures. The same set of Gabor filters, which might not be appropriate, is applied at different locations on the face. A genetic algorithm has also been used to select Gabor fea- tures for pixel classification (Campbell and Thomas, 1997) and vehicle detection (Sun et al., 2003). This basically cre- ates a population of random combinations of features, 0167-8655/$ - see front matter Ó 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2006.02.005 * Corresponding author. Tel.: +44 115 9513312; fax: +44 115 9514254. E-mail addresses: lls@cs.nott.ac.uk (L. Shen), bai@cs.nott.ac.uk (L. Bai). www.elsevier.com/locate/patrec Pattern Recognition Letters 27 (2006) 1758–1767