Two-Stage Robust Hierarchical Bayesian Collaborative Representation Classification with Deferred Spatial Refinement for Occluded Face Recognition

Face recognition under real-world occlusion remains a challenging task.A widely adopted approach is representation-based classification. This approach suppresses occlusion by assigning per-pixel weights to the coding residual.Regularized Robust Coding and Structured RRC (SRRC), two leading methods in this family, use logistic weights for this purpose.However, they share one inherent limitation: the logistic weights are near-binary and lack smooth gradation.SRRC additionally suffers from a second limitation: premature spatial coupling distorts boundary-pixel weight estimation.To address these issues, we propose Two-Stage Robust Hierarchical Bayesian Collaborative Representation Classification (TS-RHBCRC).Stage~1 refines Gamma-prior weights via maximum a posteriori coordinate ascent without spatial constraints.Consequently, the weights vary smoothly and continuously.Stage~2 alternates median filtering with weight re-estimation, thereby correcting isolated false positives and false negatives.By decoupling estimation from refinement, this deferred design substantially reduces coupling-induced distortion.On AR sunglass occlusion, TS-RHBCRC achieves 95.67%.On Yale~B, TS-RHBCRC achieves 89.14% on difficult shadow and retains 86.84% under 80% cat mask occlusion. Meanwhile, running times remain competitive.