结合注意力与局部特征融合的行人重识别算法

行人重识别是指从一堆候选图片中找到与目标最相似的行人图片，本质上是一个图像检索的子问题。为了进一步增强网络提取关键特征的能力以及抑制噪声的干扰，通过对基于注意力机制和局部特征的行人重识别算法的研究，提出了结合注意力与局部特征融合的行人重识别算法。该算法将ResNeSt-50作为骨干网络，联合软注意力与非局部注意力机制，采用双流结构分别提取行人细粒度全局特征和细粒度局部特征，通过关注不同特征之间共享的空间域信息以及同一特征不同水平区域的潜在语义相关性，创建了空间感知特征融合模块（spatial-aware feature fusion module）以及跨区域特征融合模块（cross-region feature fusion module）。在Market-1501、DukeMTMC-reID以及CUHK03数据集上的实验结果表明该算法极大程度上提升了网络的检索能力，同时与现有算法进行比较，凸显出优越性能。

Person Re-Identification Algorithm Combining Attention and Local Feature Fusion

Person re-identification refers to finding the pedestrian image most similar to the target from a bunch of candidate pictures, which is a sub-problem of image retrieval. In order to further enhance feature extraction capabilities and suppress noise interference, through studying the person re-recognition algorithm based on attention mechanism and local features, this paper proposes a person re-recognition network combing attention and local features fusion. The network adopts ResNeSt-50 as the backbone network, and uses a dual-stream structure to extract fine-grained global features and fine-grained local features separately. By focusing on the spatial domain information shared by different features and the feature correlation between different horizontal parts, the spatial-aware feature fusion module and cross-region feature fusion module are proposed. On the Market-1501, DukeMTMC-reID and CUHK03 datasets, experimental results verify that the proposed network significantly improves the retrieval capabilities. And it has a superior performance in comparison with existing algorithms.