Reconstruction of occluded facial images using asymmetrical Principal Component Analysis

When face images are occluded it becomes difficult or impossible to make a correct recognition based on only the non-occluded parts. With asymmetrical Principal Component Analysis (aPCA) it is possible to estimate occluded regions based on the content of the non-occluded pixels. Entire faces can be reconstructed from images where some parts are occluded with the occluded region being estimated based on the relation to the non-occluded region. We show the result with 3 different occlusion cases and compare two reconstruction methods. The results are measured both objectively and subjectively and it is shown that a clear and sharp image part should be used even if the resulting image have edges in it.