Learning from Examples to Generalize over Pose and Illumination

We present a neural system that recognizes faces under strong variations in pose and illumination. The generalization is learnt completely on the basis of examples of a subset of persons (the model database) in frontal and rotated view and under different illuminations. Similarities in identical pose/illumination are calculated by bunch graph matching, identity is coded by similarity rank lists. A neural network based on spike timing decodes these rank lists. We show that identity decisions can be made on the basis of few spikes. Recognition results on a large database of Chinese faces show that the transformations were successfully learnt.

• Type: Conference
• Year: 2009
• Where: ICANN
• Authors: Marco K. Müller, Rolf P. Würtz
• Keywords: ICANN 2009, Rank Lists, Similarity Rank Lists, Strong Variations