PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
March 13, 2026Journal of Image and Graphics0 citationsOpen Access

显著性引导擦除与门控正则化的遮挡行人重识别

YXYe XueyiXZXue ZhiquanJDJiang Deqi

Key Points

  • This research aims to improve the feature discrimination and robustness in occluded pedestrian re-identification.
  • Developed a multi-view learning framework integrating channel attention, significance guided erasure, and gated regularization.
  • Utilized squeeze-and-excitation residual network (SE-ResNet) to adaptively enhance discriminative features.
  • Implemented a Top-k significance guided soft erasure module to locate key areas and reduce overfitting.
  • Applied gating regularization to impose entropy constraints on feature channel selection.
  • Achieved a mean average precision (mAP) of 83.4% and Rank-1 accuracy of 92.8% on the Occluded-REID dataset.
  • On the MSMT17 dataset, mAP reached 71.4% with Rank-1 accuracy of 89.6%, showing significant improvement over existing methods.
  • Demonstrated strong performance on both synthetic and natural occlusion datasets, verifying the universality of the approach.

Abstract

目的遮挡行人重识别因可见区域缺失和遮挡噪声干扰,导致特征判别性不足、鲁棒性差。现有方法多局限于单视图表征,存在通道判别性建模不充分、显著区域过拟合等问题。方法提出一种融合通道注意力增强、显著性引导擦除与门控正则化的多视角学习框架。该方法首先采用挤压激励残差网络(squeeze-and-excitation residual network,SE-ResNet)作为骨干网络,通过通道注意力机制自适应增强判别性特征。进而,设计Top-k显著性引导软擦除模块,利用类激活图定位关键区域并进行平滑抑制,迫使网络发掘多样化辅助特征,以缓解过拟合。此外,引入门控正则化机制,对特征通道选择概率施加熵约束,提升选择过程的稀疏性与稳定性。在训练阶段,通过加权融合多视图特征并借助特征对齐损失实现信息协同;推理阶段仅需单视图特征即可完成高效检索。结果在四个代表性数据集上的实验表明,本方法均取得领先性能。其中,在专门针对遮挡行人重识别的数据集Occluded-REID上,平均精度均值(mean average precision,mAP)与首位命中率(Rank-1 accuracy,Rank-1)分别达到83.4%与92.8%;在大型多场景复杂数据集MSMT17上,分别达到71.4%与89.6%,较先进方法MVIIP有显著提升。在基于DukeMTMC构建的合成遮挡数据集P-DukeMTMC和经典通用行人重识别基准数据集Market-1501上,本方法同样表现优异,验证了其针对合成与自然遮挡的普适性。结论本方法在遮挡与多场景任务中均表现出良好的适应性、鲁棒性和泛化能力,适用于复杂监控场景下的行人检索。

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Xueyi et al. (2026) studied this question.

synapsesocial.com/papers/69b3aad702a1e69014ccb904https://doi.org/10.11834/jig.250576
Ask AI
Helpful
Bookmark
Share
View Full Paper