Dual-subspaces based quantitative analysis of facial appearance

We propose a subspace based method for quantitative analysis of facial appearance. We first collect both natural and cosmetic female facial images and then construct a subspace of female facial images using principal component analysis (PCA). We divide the subspace into two subspaces: one is global subspace which is constructed by eigenvectors with larger eigenvalues and another one is local subspace which is constructed by eigenvectors with smaller eigenvalues. Both natural and cosmetic facial images are projected to global subspace and local subspace, respectively. The difference (distance) of the projection between the natural facial image and cosmetic facial image is used as a quantitative measure of make-up effect. We found that the difference (distance) in global subspace represents skin color information and the difference (distance) in local subspace represents skin texture information. The quantitative analysis results are strongly correlated with the results of psychological test.