Reconstruction of Partially Occluded Face by Fast Recursive PCA

This paper proposes a fast recursive PCA (principal component analysis) algorithm to remove face occlusions. In training phase, all faces are normalized by two eye centers and two mouth corners, and eigenvectors (eigenfaces) were obtained by PCA analysis. In test phase, face occlusion is removed by iteratively perform two steps of analysis and synthesis. New damaged face is first normalized by clicking four feature points, and PCA coefficients are obtained in analysis step. In synthesis step, reconstructed face is obtained by linear combining eigenfaces, and coefficients error between two successive analyses is used for fast PCA compensation. Experimental results on training and test faces show that the proposed algorithm convergences faster than classical PCA compensation and reconstructed faces are natural.