Abstract: Currently, unsupervised generalization-based person re-identification algorithms may overlook the fine-grained feature attention in certain situations where background information is needed to assist in judgment, leading to the filtration of background information and a decrease in recognition accuracy. To address this issue, a dynamic query-aware algorithm based on sparse attention is proposed. Firstly, the squeeze and concat (SPC) module is utilized to obtain feature maps with different channel numbers. Secondly, a bi-level routing attention mechanism is employed to extract attention weights between feature maps of different scales, resulting in hierarchical channel attention vectors. Then, the weights of the hierarchical channel attention vectors are recalibrated. Finally, the recalibrated weights are used to weight the corresponding feature maps, thereby generating multi-scale feature maps with richer and more refined feature information. Experimental results on large-scale public datasets (Market-1501, DukeMTMC-reID, MSMT17) demonstrate that the proposed model achieves improvements of 3.2, 4.4, and 15.4 percentage points in Rank-1 accuracy and 5.5, 8.3, and 16.2 percentage points in mAP, respectively. In comparison with existing state-of-the-art algorithms, this model enables better interaction between local and global feature channels, enhancing the model’s ability to perceive detailed image features.

Key words: person re-identification, fine-grained features, sparse attention mechanism, dynamic query-aware, feature recombination

摘要： 目前无监督的泛化行人重识别算法在某些需要背景信息辅助判断的情况下，可能会忽视图像的局部区域对细粒度特征的关注，导致背景信息被过滤掉，从而降低识别精度。针对上述问题，提出了一种基于稀疏注意力的动态查询感知算法。通过挤压拼接（squeeze and concat，SPC）模块，获取不同通道数的特征图。利用双层路由感知注意力机制，提取不同尺度特征图之间的注意力权重，得到逐级通道注意力向量。对逐级通道注意力向量的权重进行重新校准。将重新标定的权重与相应的特征图进行加权，输出具有更丰富细化特征信息的多尺度特征图。所提模型在大规模公开数据集（Market-1501、DukeMTMC-reID、MSMT17）上进行实验，相较于基线模型Rank-1分别提高了3.2、4.4、15.4个百分点，mAP分别提高了5.5、8.3、16.2个百分点，与现有前沿算法相比，能够实现更好的局部和全局特征通道之间的信息交互，提升模型对图像特征的细节感知能力。

关键词: 行人重识别, 细粒度特征, 稀疏注意力机制, 动态查询感知, 特征重组