范馨月,张阔,张干,李嘉辉.细微特征增强的多级联合聚类跨模态行人重识别算法[J].电子测量与仪器学报,2024,38(3):94-103 |
细微特征增强的多级联合聚类跨模态行人重识别算法 |
Cross-modal person re-identification algorithm based on multi-level joinclustering with subtle feature enhancement |
|
DOI: |
中文关键词: 行人重识别 跨模态 随机颜色转换 细微特征增强 多级联合聚类学习 |
英文关键词:person re-identification cross-modality random color transition subtle feature enhancement multilevel joint clustering learning |
基金项目:国家自然科学基金(62271096)项目资助 |
|
Author | Institution |
Fan Xinyue | School of Communication and Information Engineering, Chongqing University of Posts
and Telecommunications, Chongqing 400065, China |
Zhang Kuo | School of Communication and Information Engineering, Chongqing University of Posts
and Telecommunications, Chongqing 400065, China |
Zhang Gan | School of Communication and Information Engineering, Chongqing University of Posts
and Telecommunications, Chongqing 400065, China |
Li Jiahui | School of Communication and Information Engineering, Chongqing University of Posts
and Telecommunications, Chongqing 400065, China |
|
摘要点击次数: 319 |
全文下载次数: 2057 |
中文摘要: |
目前跨模态行人重识别研究注重于通过身份标签在全局特征或局部特征上提取模态共享特征来减少模态差异,但却忽视了具有辨别力的细微特征。为此提出了一种基于特征增强的聚类学习网络,该网络通过全局和局部特征来挖掘并增强不同模态的细微特征,并结合多级联合聚类学习策略,最小化模态差异和类内变化。针对训练数据设计了随机颜色转换模块,在图像输入端增加模态之间的交互,以克服颜色偏差的影响。通过在公共数据集上进行实验,验证了所提方法的有效性,其中在SYSU-MM01数据集的全搜索模式下Rank-1和mAP分别达到了70.52%和64.02%;在RegDB数据集的V2I检索模式下Rank-1和mAP分别达到了88.88%和80.93%。 |
英文摘要: |
The current cross-modal person re-identification research focuses on extracting modality-shared features from global features or local features via identity labels to reduce modality differences, but ignores the Subtle features of discernment. This paper proposes a feature enhanced clustering learning (FECL) network. The network mines and enhances the subtle features of different modalities through global and local features, and combines a multilevel joint clustering learning strategy to minimize the modal differences and intraclass variation. In addition, this paper also designs a random color transition module for training data, which increases the interaction between modalities at the image input to overcome the influence of color deviation. The experiments on public datasets verify the effectiveness of the proposed methods. In the All search mode of SYSU-MM01 dataset, the Rank-1 and mAP reach 70.52% and 64.02%. In the V2I retrieval mode of RegDB dataset, the Rank-1 and mAP reach 88.88% and 80.93%. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|