Occlusion-invariant face recognition using simultaneous segmentation

When using convolutional neural network (CNN) models to extract features of an occluded face, the occluded part will inevitably be embedded into the representation just as with other facial regions. Existing methods deal with occluded face recognition either by augmenting the training dataset with synthesized occluded faces or by segmenting occlusions first and subsequently recognize the face based on unoccluded facial regions. Instead, simultaneous occlusion segmentation and face recognition is developed to make the most of these correlated two tasks. This is inspired by the phenomenon that features corrupted by occlusion are traceable within a CNN trained to segment occluded parts in face images. Specifically, a simultaneous occlusion invariant deep network (SOIDN) is proposed that contains simultaneously operating face recognition and occlusion segmentation networks coupled with an occlusion mask adaptor module as their bridge to learn occlusion invariant features. The training of SOIDN is jointly supervised by classification and segmentation losses aiming to obtain (1) occlusion invariant features, (2) occlusion segmentation, and (3) an occlusion feature mask that weighs the reliability of features. Experiments on synthesized occluded dataset (e.g. LFW-occ) and real occluded face dataset (e.g. AR) demonstrate that SOIDN outperforms state of the art methods for face verification and identification.