Modeling face appearance with nonlinear independent component analysis

Appearance-based approach is one of popular methods for face analysis. How to describe face appearance is a key issue for appearance based face analysis. Principal component analysis (PCA) and independent component analysis (ICA) are two successful and well-studied linear unsupervised representation methods of face appearance. However, there exist complicate nonlinear variations in real face images due to pose, illumination, expression variations and so on, so it is inadequate for PCA and ICA to describe these nonlinear relations in real face images because of their linear properties in nature. In this paper, a nonlinear ICA is proposed to model face appearance, which combines the nonlinear kernel trick with ICA. First, the kernel trick is employed to project the input image data into a high-dimensional implicit feature space F with a nonlinear mapping, and then ICA is performed in F to produce nonlinear independent components of input data. We call it kernel ICA or KICA. In the experiments, the polynomial kernel is used, and experimental results show the proposed method has an encouraging performance.