基于局部特征关联与全局注意力机制的行人重识别

行人重识别的难点在于行人之间的结构信息差异较小导致特征难以区分。结合全局关系注意力机制与局部特征关联方法提出一种改进的特征关联算法。通过水平切分全局注意力机制的特征得到多个局部特征，并进行逐个关联识别，利用局部特征关联与全局语义信息提取关键特征信息。在此基础上，采用交叉熵与三元组损失函数训练处理后的局部特征。在CUHK03-Labeled、Market1501、DukeMTMC-reID数据集上的实验结果表明，该算法首位命中率分别为81.6%、95.6%、89.5%，相比GCP、MGN、BAS-reID等算法具有更强的识别能力与自适应性。

Person Re-Identification Based on Local Feature Relation and Global Attention Mechanism

In person re-identification, the difference of structural information between pedestrians is minor, making it hard to distinguish between the pedestrians with similar overall features.To address the problem, an improved feature relation algorithm is proposed by combining global relation attention mechanism and local feature relations.The algorithm horizontally segments the features of the global attention mechanism to obtain multiple local features, which are subsequently related and recognized.Then the local feature relations and the global semantic information are used to extract key feature information.On this basis, the cross entropy and the triplet loss function are used to train the processed local features.The experimental results show that the proposed algorithm displays a first place accuracy of 81.6% on CUHK03-Labeled, 95.6% on Market1501, and 89.5% on DukeMTMC-reID, providing high recognition performance and adaptability.