Robust Face Recognition in Rotated Eigen Space
S. Chen, Tianlong Shan, Brian C. Lovell · Image and Vision Computing New Zealand · 2007
Face recognition is a very complex classification problem due to nuisance variations in different conditions. Most face recognition approaches either assume constant lighting condition or standard facial expressions, thus cannot deal with both kinds of variations simultaneously. Principal Component Analysis (PCA) cannot handle complex pattern variations such as illumination and expression. Adaptive PCA rotates eigenspace to extract more representative features thus improving the performance. In this paper, we present a way to extract various sets of features by different eigenspace rotations and propose a method to fuse these features to generate nonorthogonal mappings for face recognition. The proposed method is tested on the Asian Face Database with 856 images from 107 subjects with 5 lighting conditions and 4 expressions. We register only one normally lit neutral face image and test on the remaining face images with variations. Experiments show a 95% classification accuracy and a 20% reduction in error rate. This illustrates that the fused features can provide significantly improved pattern classification.