Face Recognition Using Recurrent High-Order Associative Memories Iulian B. Ciocoiu Technical University of Iasi, Romania Faculty of Electronics and Telecommunications iciocoiu@etc.tuiasi.ro Abstract. A novel face recognition approach is proposed, based on the use of compressed discriminative features and recurrent neural classifiers. Low-dimensional feature vectors are extracted through a combined effect of wavelet decomposition and subspace projections. The classifier is implemented as a special gradient-type recurrent analog neural network acting as an associative memory. The system exhibits stable equilibrium points in predefined positions given by the feature vectors of the training set. Experimental results for the Olivetti database are reported, indicating improved performances over standard PCA and LDA-based face recognition approaches. 1. Introduction Face recognition has represented for more than one decade one of the most active research areas in pattern recognition. A plethora of approaches have been proposed and evaluation standards have been defined, but current state-of-the-art solutions still need to be improved in order to cope with the recognition rates and robustness requirements of commercial products. Most of the approaches may be classified into two categories [2]: a) geometric feature-based techniques, relying on the identification of specific components of a face such as eyes, nose, mouth, and distances among them b) holistic template-based techniques, usually based on projecting the original (high- dimensional) images onto lower dimensional subspaces spanned by specific basis vectors. Eigenfaces [13] represent a de facto standard for the second approach and, although superior solutions exist, still defines a performance reference against which any new method is compared. Face recognition systems usually include three modules, i.e. the preprocessing stage, feature extraction, and classification. Although the novelty aspect of the present paper is mainly related to the classifier, we present key elements of the other components in the following: • similar to other approaches [3, 14], we perform a multiresolution decomposition of the original images based on the Discrete Wavelet Transform (DWT) and keep only the low-frequency components (Figure 1). Besides dimensionality reduction this procedure is also known to offer face expression invariance. ESANN'2004 proceedings - European Symposium on Artificial Neural Networks Bruges (Belgium), 28-30 April 2004, d-side publi., ISBN 2-930307-04-8, pp. 567-572