基于非对称增强注意力与特征交叉融合的行人重识别方法

Pedestrian Re-identification Method Based on Asymmetric Enhanced Attention and Feature Cross Fusion

针对现有的行人重识别方法提取到的特征信息充分性与辨识性不足导致检索精度低的问题,提出一种基于非对称增强注意力与特征交叉融合的行人重识别方法。首先,构建非对称增强注意力模块,通过多重池化聚合的跨邻域通道交互注意力增强显著特征表示,使网络聚焦于图像中的行人区域;其次,考虑到网络各层特征间的差异性与关联性,构建特征交叉融合模块,利用交叉融合方式实现同层不同级特征的跨层级融合,进而实现多尺度融合;最后,水平切分输出特征以获取局部特征,从而实现在特定区域上描述行人。在Market1501、DukeMTMC-reID与CUHK03这3个公开数据集上对提出的方法进行了验证,首位命中率(Rank-1)分别达到了93.5%、85.1%和64.3%,证明了该方法在提升行人重识别性能上具有优越性。

In order to solve the problem of poor retrieval accuracy due to the lack of sufficient and recognizable feature information extracted by the existing pedestrian re-identification methods,a pedestrian re-identification method based on the asymmetric enhanced attention and feature cross fusion is proposed.First,an asymmetric enhanced attention module is constructed,and the salient feature representation is enhanced through the cross-neighbor channel interactive attention that is aggregated by multiple pooling,so that the pedestrian area in the image is focused by the network.Then,taking into account the differences and relevance between the features of each layer of the network,a feature cross fusion module is constructed,and the cross fusion method is used to achieve cross-level fusion of features at different levels in the same layer,thereby realizing multi-scale fusion.Finally,the output features are segmented horizontally to obtain local features,so as to describe pedestrians in a specific area.The proposed method is verified on the three public data sets of Market1501,DukeMTMC-reID and CUHK03.Rank-1 reach 93.5%,85.1%and 64.3%respectively,proving that the method is superior in improving the performance of pedestrian re-identification.