多分支融合注意力机制的行人重识别方法

Person re-identification method based on multi-branch fusion attention mechanism

针对基于局部特征的行人重识别方法在存在背景干扰时鲁棒性较差的问题,提出一种多分支融合注意力机制的行人重识别方法。该方法包含两个分支网络,其中全局分支通过融合注意力机制获取更具显著性的全局特征信息;局部分支通过提取多尺度的语义信息和不同粒度的特征挖掘行人非显著性的局部特征,增强特征的丰富性,引入特征补充避免硬分割造成图像分割边缘信息丢失。训练阶段使用softmax损失、三元组损失和中心损失联合优化网络,在Market1501、DukeMTMC-reID和CUHK03数据集上进行实验对比分析,其结果表明,该方法能够有效提高行人重识别的准确率。

To solve the problem that person re-identification methods based on local feature have poor robustness in the presence of background interference,a person re-identification method based on multi-branch fusion attention mechanism was presented.The method consisted of two branch networks,in which the global branch obtained more significant global feature information by fusing the attention mechanism.The local branch mined nonsignificant local features of pedestrians by extracting multi-scale semantic information and different granularity features to enhance the richness of the features,and feature supplement was introduced to avoid the loss of edge information caused by hard segmentation.Softmax loss,triplet loss and center loss were used to optimize the network jointly during the training phase.Results of experiments on Market1501,DukeMTMC-reID and CUHK03 datasets show that the method can effectively improve the accuracy of person re-identification.