Deferred Spatial Refinement for Robust Hierarchical Bayesian Face Recognition under Occlusion

Face recognition under occlusion is still challenging.Robust representation-based classification reduces occlusion effects by weighting residual pixels.A widely used pixel weight is the logistic function, as adopted by Regularized Robust Coding (RRC).However, logistic weights collapse to near-binary values, easily producing isolated false positives (FPs) and false negatives (FNs).To correct these isolated errors, Structured RRC (SRRC) introduces spatial constraints that couple the weights of neighboring pixels.Yet, the premature spatial coupling impairs the estimation of boundary pixel weights, limiting the gain over RRC.To address these issues, we propose a Two-Stage Robust Hierarchical Bayesian Collaborative Representation Classification (TS-RHBCRC) method.Stage~1 refines polynomial-tailed Gamma-prior weights via maximum a posteriori coordinate ascent without spatial constraints.Stage~2 applies deferred spatial refinement through median filtering and weight re-estimation.The decoupled design enables gradual weight gradation, corrects most isolated FPs and FNs, and avoids coupling-induced distortion.Experiments on the AR and Extended Yale~B databases demonstrate the effectiveness, robustness, and efficiency of TS-RHBCRC under occlusions.