ICA-based Neighborhood Preserving Analysis for Face Recognition
Haifeng Hu
DOI: https://doi.org/10.1016/j.cviu.2008.05.003
IF: 4.886
2008-01-01
Computer Vision and Image Understanding
Abstract:We present a new dimensionality reduction method for face recognition, which is called independent component based neighborhood preserving analysis (IC-NPA). In this paper, NPA is firstly proposed which can keep the strong discriminating power of LDA while preserving the intrinsic geometry of the in-class data samples. As NPA depends on the second-order statistical structure between pixels in the face images, it cannot find the important information contained in the high-order relationships among the image pixels. Therefore, we propose IC-NPA method which combines ICA and NPA. In this method, NPA is performed on the reduced ICA subspace which is constructed by the statistically independent components of face images. IC-NPA can fully consider the statistical property of the input feature. Furthermore, it can find an embedding that preserves local information. In this way, IC-NPA shows more discriminating power than the traditional subspace methods when dealing with the variations resulting from changes in lighting, facial expression, and pose. The feasibility of the proposed method has been successfully tested on both frontal and pose-angled face recognition, using two data sets from the FERET database and the CAS-PEAL database, respectively. The experiment results indicate that the IC-NPA shows better performance than the popular method, such as the Eigenface method, the ICA method, the LDA-based method and the Laplacianface method.