The Impact of ARTMAP to Appearance-based Face Verification

We propose a novel approach for matching and preservation of face images based on Adaptive Resonance Theory MAPping (ARTMAP) network. ART networks possess incrementally growing structure and provide stable on-line learning, which ensures that all pat:ems presented to the network, will be learned and compactly stored. Moreover the network's weights will be adapted after each classification. These characteristics are important for successful recognition of an object, which patterns are quite changeable in time. In our implementation called FaceART the network is learned from raw images as well as from eigenfaces decomposition coefficients. In order to compare the error rates of the implemented system to existing academic face recognition systems the XM2VTS database with Lausanne protocol is employed. We show that compared to the nearest neighbor rule the presented classification approach has better verification performance and more compact template representation.