基于多尺度混合注意力的行人重识别模型

Multi-scale hybrid attention network for person re-identification

针对目前行人重识别中复杂背景下人物容易被遮挡物掩盖和人物特征不明显导致模型难以提取重点特征的问题,提出一种基于多尺度混合注意力的行人重识别模型。在特征提取部分,设计一种特征提取模块,即多尺度混合注意力残差块,通过该模块可得到丰富上下文关系,在频域范围内获得更丰富的通道特征信息。通过在大型数据集Market1501、DukeMTMC-reID、CUHK03-L、CUHK03-D上的验证,行人重识别的精度得到了有效提升,与目前先进的模型结果对比,mAP精度提升了0.3%、2.1%、0.7%、2.9%,Rank1在DukeMTMC-reID、CUHK03-L数据集提升了0.2%、0.8%。

A pedestrian re-identification model based on multi-scale mixed attention was proposed to address the challenges of occlusion and unclear pedestrian features in complex backgrounds.A feature extraction module called multi-scale mixed attention residual block was designed to capture rich contextual relationships and obtain more comprehensive channel feature information in the frequency domain.Through experiments on large-scale datasets including Market1501,DukeMTMC-reID,CUHK03-L and CUHK03-D,the proposed model demonstrates effective improvements in pedestrian re-identification accuracy.When compared with state-of-the-art models,the mAP accuracy is improved by 0.3%,2.1%,0.7%,and 2.9%respectively,while the Rank-1 accuracy is improved by 0.2%and 0.8%on the DukeMTMC-reID and CUHK03-L datasets,respectively.