Fusion of ICA Spatial, Temporal and Localized Features for Face Recognition

Independent component analysis (ICA) has found its application in face recognition successfully. In practice several ICA representations can be derived. Particularly they include spatial ICA, spatiotemporal ICA, and localized spatiotemporal ICA, which respectively extract features of face images in terms of space domain, time-space domain, and local region. Our work has shown that while spatiotemporal ICA outperforms other ICA representations, further improvement can be made by a fusion of variety of ICA features. However, simply combining all features will not work as well as expected. For this reason an optimization method for feature selection and combination is proposed in this paper. We present here an optimizing process of feature selection about which features and how many features from each individual ICA feature set are selected. The experimental results show that feature fusion method can improve face recognition rate up to 94.62% compared with that of 86.43% by using spatiotemporal ICA alone.