Reconstruction of occluded facial images using asymmetrical Principal Component Analysis

When only non-occluded image parts are available for facial images it is difficult or impossible to correctly recognize the person in the image. The problem addressed in this work is reconstruction of the occluded parts in facial images; e. g. eyes covered with sunglasses. Asymmetrical Principal Component Analysis (aPCA) allows estimation of occluded facial parts based on the content of the facial parts which are visible. aPCA is used to estimate full non-occluded faces from 3 kinds of occlusion with 2 different reconstruction methods in this work and we present the results with both objective and subjective evaluation. The subjective evaluation shows that clear and sharp image regions are preferred even if this results in visible edges in the images. The method also performs well when a different facial expression than the one in the database is used to calculate the reconstruction parameters.