Manifold synthesis: Predicting a manifold from one sample
Li-Yun Lo, Ju-Chin Chen · 2011
This study proposes a manifold synthesis approach based on the regression model to predict a manifold from one sample in order to cope with the single sample per person problem for face recognition. To reduce the dimensionality and preserve the data relation in the input image space, local preserving projection (LPP) is applied. Thus for each subject the facial images with pose angles can be represented by a manifold in the LPP space but the pose and the subject's appearance information are depicted in the same manifold. Hence, the manifold alignment approach is applied to extract the identity-invariant manifold that only the pose information is included in the joint latent space. Based on the regression model, two approaches are developed in the LPP and joint latent space to estimate the manifold when one frontal image is given. In the experimental results, FacePix database is conducted to evaluate the system performance.