Occluded person re-identification (Re-ID) requires matching identities when body cues are missing or obscured by unrelated objects. Pose and viewpoint variations further disrupt semantic correspondence between partially visible persons. We propose the Progressive Hierarchical Collaborative Perception Network (PHCP-Net), which integrates diffusion-guided reconstruction, hierarchical collaboration, and feature enhancement. A frozen pose-guided diffusion model generates reconstructed counterparts that provide complementary body cues during training. Within each branch, the Collaborative Interaction Module (CIM) recursively integrates class-token features from shallow to deep stages. The Calibration Optimization Unit (COU) uses reconstructed features as semantic anchors to calibrate intermediate features in the original branch. The Enhanced Fusion Module (EFM) combines the final Vision Transformer (ViT) token-level feature with the CIM output to produce an enhanced identity embedding. The Distribution Constraint Module (DCM) then aligns the two branches through probabilistic, statistical, and identity-conditioned constraints. Only the original branch, comprising ViT, CIM, and EFM, is retained for inference. PHCP-Net achieves 76.20% Rank-1 accuracy and 66.50% mAP on Occluded-Duke, yielding the highest Rank-1 accuracy and competitive mAP among the compared methods. On Occluded-REID, it achieves the highest compared Rank-1 accuracy and mAP, at 90.60% and 84.50%, respectively. Ablation studies support the complementary contributions of hierarchical fusion, intermediate calibration, and final representation alignment to occlusion-robust identity learning.
The integration of rapidly developing social networks into large-scale group decision-making has attracted growing attention, particularly regarding consensus-reaching process costs. Traditional minimum cost consensus models (MCCMs) focus primarily on minimizing total consensus cost, often implicitly assuming unconditional acceptance of opinion adjustments by decision makers (DMs) and neglecting fairness in adjustment allocation. To overcome these limitations, this paper proposes a dual-perspective two-stage robust MCCM. From the DMs’ perspective, fair opinion adjustment schemes are derived via cooperative game theory in the first stage. From the moderator’s perspective, the second stage determines the optimal consensus cost based on the first stage’s adjustment benchmarks, while accounting for heterogeneous risk preferences shaped by DMs’ structural positions in the trust-based network. A Gini index-based fairness measure is further introduced to ensure the second stage’s fairness level is no lower than that of the first stage. Moreover, a hybrid weighting scheme and an improved community detection algorithm are developed to jointly capture opinion deviation and trust relationships in social networks. To address uncertainty in unit adjustment costs, both classical and data-driven robust optimization approaches are incorporated, enabling the proposed framework to adapt to diverse decision-making environments. Finally, numerical experiments demonstrate the effectiveness, robustness, and fairness of the proposed approach.
