Recovery of occluded face using direct combined model-based particle filter

In this paper, we present a Bayesian framework for recovering the occluded facial image without the aid of manual face alignment or user-specified occlusion region. The proposed Bayesian framework unifies the recovery stage with face alignment and occlusion detection, and such complex probability distribution is represented by a particle set. Into this framework, each particle is one possible solution of the recovered face which is composed of several patches. First, the occluded facial patches of each particle are detected, and then are recovered by inferring their local facial details from other non-occluded patches. Further, by including the global facial geometry as a constraint, the recovered results are robust to the local image noise which then cause the alignment parameters are accurately calculated. Particularly, we also propose a novel direct combined model (DCM)-based particle filter that utilizes the face specific prior knowledge to perform such particle-based solution efficiently and robustly. Our extensive experiment results demonstrate that the recovered images are quantitatively closer to the ground truth without manual involvement.