全局注意力机制与关系网络驱动的行人重识别

Person re-identification driven by global attention mechanism and relation network

针对行人重识别(re-identification,ReID)中因遮挡、姿态变化导致模型提取关键特征能力不足的问题,提出一种全局注意力机制与关系网络驱动的行人重识别算法。首先,在主干网络部分,将全局注意力模块嵌入ResNet-50网络中,捕获空间维度和通道维度的权重信息;其次,在关系网络部分,通过水平切分得到不同尺度的多个局部特征,并设计了2个模块分别用于提取全局对比特征及局部特征间的关系特征,并将局部关系特征与全局对比信息融合后送到分类网络;最后,在损失优化阶段,采用Circle Loss、三元组损失与交叉熵损失的联合损失进行网络训练。在5个常用数据集上进行实验并对实验结果进行分析,其中Occluded-DukeMTMC和Market1501数据集上rank-1值分别达到63.2%和95.4%,mAP(mean average precision)值分别为53.8%和88.2%,体现了所提算法的先进性。

To address the issue of the insufficient ability of existing models in the extraction of crucial features in person re-identification(ReID)caused by occlusions and pose variations,a person re-identification algorithm driven by a global attention mechanism and relation network is proposed.Firstly,in the backbone network,a global attention module is embedded into the ResNet-50 network to capture weight information in both spatial and channel dimensions.Secondly,in the relation network,multiple local features of different scales are obtained through horizontal partitioning.Two modules are designed to extract the global contrast features and relation features among local features.The local relation features are fused with global contrast information and fed into the classification network.Finally,in the loss optimization stage,a joint loss consisting of Circle Loss,triplet loss,and cross-entropy loss is employed for network training.Experiments are conducted on five commonly used datasets,and the results are analyzed.The proposed algorithm achieves a rank-1 accuracy of 63.2%and 95.4%,as well as mAP scores of 53.8%and 88.2%on the Occluded-DukeMTMC and Market1501 datasets,respectively,demonstrating its advancement.