Region Selection Enhancement Joint Parallel Decoder for Person Re-Identification

To mitigate the cluttered information caused by occlusion interference and background noise in person re-identification (Re-ID) tasks, as well as to alleviate the matching interference induced by diverse pedestrian postures, this paper proposes a novel Region Selection Enhancement Joint Parallel Decoder (RSEPD) framework for person Re-ID. The proposed region selection enhancement technique generates feature maps containing noise information by performing data augmentation within constrained regions, encouraging the model to disregard noise introduced by the background and occlusion while reducing the impact of complex postures. Additionally, a novel dual-branch decoder are designed to process multiple feature maps in a parallel and independent manner. A two-stage penalty strategy is further employed to guide the convergence of both the feature extractor and the decoder. Finally, experimental results demonstrate that the proposed method achieves 91.1% Rank-1 accuracy and 79.4% mAP on the DukeMTMC-reID dataset, effectively mitigating the interference caused by occlusion, background noise, and posture misalignment. Moreover, compared to several recent approaches, the proposed method exhibits strong competitive performance.