%0 Journal Article
%T 基于多粒度注意力的可见光–红外行人重识别模态特异性算法<br>Multi-Granularity Modality Specific Attention Algorithm for Visible-Infrared Person Re-Identification
%A 王娅迪
%J Computer Science and Application
%P 83-92
%@ 2161-881X
%D 2023
%I Hans Publishing
%R 10.12677/CSA.2023.131009
%X <span style=\"font-size:14px;\">可见光–红外行人重识别(Visible-infrared Person Re-identification, VI Re-ID)旨在根据裁剪后的查询行人样本，检索另一模态的画廊集中相同身份的行人样本。现有基于特征融合的方法主要采用双流特征提取模型来解决此任务，但由于两种模态的数据来源存在差异，不同模态特征提取分支存在特征粒度不一致的问题。为了解决这一问题，本文提出一种多粒度空间注意力模块，并进一步提出模态特异性双注意力模块，使用分散的细粒度注意力提取包含更多细节信息的可见光样本注意力描述符，使用集中的粗粒度注意力提取包含更多模糊信息的红外样本注意力描述符。实验证明融入多粒度方法的注意力模块有效的提高了双流特征提取模型在SYSU-MM01和RegDB数据集上的精度。</span><br />
<span style=\"font-size:14px;\">Visible-infrared Person Re-identification aims to recognize pedestrian images with the same identity as query pedestrian image in the gallery set of another modality. Existing methods based on two-stream feature extraction network have been widely used to solve this task. However, due to the different data sources, the feature granularities in different modality feature extraction branches are inconsistent. To alleviate this issue, we propose a multi-granularity spatial attention model and further proposes a modality specific attention module, which use decentralized fine-grained attention to extract the attention descriptor of visible samples containing more detailed information, and use concentrated coarse-grained attention to extract the attention descriptor of infrared samples containing more fuzzy information. Extensive experiments show that our method greatly improves the performance of the end-to-end framework on two widely used visible-infrared Person Re-identification datasets, SYSU-MM01 and RegDB.</span>
%K 跨模态行人重识别，多粒度，注意力机制，深度学习，计算机视觉<br>Cross-Modality Person Re-Identification
%K Multi-Granularity
%K Attention Mechanism
%K Deep Learning
%K Computer Vision
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=60461