| 宋婉茹,郝川艳,郑洁莹,刘峰.多粒度共享-解离相关网络支持下的跨模态行人重识别算法[J].电子测量与仪器学报,2025,39(10):1-11 |
| 多粒度共享-解离相关网络支持下的跨模态行人重识别算法 |
| Multi-granularity shared-disentangling relation network forcross-modality person re-identification |
| |
| DOI： |
| 中文关键词: 行人重识别 可见光-红外 共享-解离 多粒度 特征学习 关系网络 |
| 英文关键词:person re-identification visible-infrared shared-disentangling multi-granularity feature learning relation network |
| 基金项目:国家自然科学基金（62307025）、江苏省高校自然科学基金面上项目（22KJB520025）资助 |
|
|
| Author | Institution |
| Song Wanru | School of Educational Science and Technology, Nanjing University of Posts and Telecommunications,
Nanjing 210046, China |
| Hao Chuanyan | School of Educational Science and Technology, Nanjing University of Posts and Telecommunications,
Nanjing 210046, China |
| Zheng Jieying | School of Communication and Information Engineering, Nanjing University of
Posts and Telecommunications, Nanjing 210023, China |
| Liu Feng | 1.School of Educational Science and Technology, Nanjing University of Posts and Telecommunications,
Nanjing 210046, China; 2.School of Communication and Information Engineering, Nanjing University of
Posts and Telecommunications, Nanjing 210023, China |
|
| 摘要点击次数: 735 |
| 全文下载次数: 710 |
| 中文摘要: |
| 随着智能安防系统的不断升级，面向全天候监控实现行人检索成为了相关领域热点之一，可见光 红外跨模态行人重识别的研究应运而生。该研究面临的主要挑战是同一行人在不同模态的图像间展现出巨大的差异。现有方法通过探索不同模态之间共享信息，来减少同一行人在两种模态下的特征差异。为了进一步提升跨模态行人重识别的准确率，提出了一种多粒度共享-解离相关网络，通过共享-解离模块的嵌入，对主干网络中参数共享分支进行复制和分解，打破了原有基准模型在多粒度特征提取上的局限；通过多粒度相关特征学习模块的设计，充分挖掘了行人跨模态不变的身体结构关联信息，优化了全局 局部特征的对齐方案；通过多层次的损失函数构建，为模型的训练提供了有效的监督，提升了模型的判别力和鲁棒性。该算法在公开数据集SYSU-MM01和RegDB上均获得优秀的性能，其中，SYSU-MM01全搜索模式下Rank-1和平均精度均值（mAP）分别达到74.70%和71.79%；在RegDB的两种检索模式下，Rank-1和mAP均高于90%，准确率优于多种先进方法。实验显示该网络在跨模态特征对齐和复杂场景适应性方面具有一定优势。 |
| 英文摘要: |
| With the continuous development of intelligent security systems, pedestrian retrieval for all-day surveillance has become one of the research hotspots. Thus, the research of visible-infrared cross-modality person re-identification has emerged. The main challenge faced in this task is the huge discrepancy between visible and infrared images of the same pedestrian. Existing methods focus on exploring the shared information and reducing the feature variances of the same pedestrian in the two modalities. To further improve the accuracy of the task, this paper proposes multi-granularity shared-disentangling relation network for re-identification. By embedding the shared-disentangling module, the parameter-sharing branch of the backbone is replicated and decomposed, thus breaking limitations of the original benchmark model in multi-granularity feature extraction. By designing the multi-granularity relation feature learning module, the modality-invariant correlation information of the pedestrian body is fully explored, enhancing the learning of the shared features. And through constructing a loss function in multiple levels, effective supervision is available for the training of the model, and the global-local feature alignment scheme is optimized. The proposed algorithm obtains superior performance on both public datasets named SYSU-MM01 and RegDB. The Rank-1 and mAP in All-search mode on the SYSU-MM01 dataset can reach 74.70% and 71.79% respectively. In both retrieval modes of RegDB, Rank-1 and mAP are higher than 90%, and the accuracy is superior to many state-of-the-art methods. Experiments demonstrate the advantages of this network in cross-modality feature alignment and complex scene adaptation. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
