| 基于多尺度加权特征融合的行人重识别方法研究 |
| |
| 引用本文： | 孙劲光,吴明岩.基于多尺度加权特征融合的行人重识别方法研究[J].信号处理,2022,38(10):2201-2210. |
| |
| 作者姓名： | 孙劲光 吴明岩 |
| |
| 作者单位： | 辽宁工程技术大学电子与信息工程学院，辽宁 葫芦岛 125000 |
| |
| 基金项目： | 国家重点研发计划2018YFB1403303辽宁省教育厅重点攻关项目LJ2020ZD005 |
| |
| 摘 要： | 针对目前由于行人重识别普遍存在的遮挡以及多姿态变化等原因，导致的行人重识别率低的问题，提出一种基于多尺度加权特征融合的行人重识别方法（Person Re-identification Method Based on Multi-scale Weighted Feature Fusion， MSWF）。该方法首先使用基准网络ResNeSt-50提取图像特征，获得下采样3倍、下采样4倍和下采样5倍的特征图，输入到加权特征金字塔网络中，然后使用快速归一化融合方法进行特征融合，在特征融合中引入加权操作可以让模型在训练过程中学习如何给融合特征的权重值进行分配，这样可以充分利用不同尺度的特征，获得更加丰富的行人特征。最后将融合后的富含语义信息的高层特征作为全局特征，将融合后的高分辨率特征作为局部特征。在训练过程中，联合Softmax分类损失函数、三元组损失函数和中心损失函数对模型进行训练，在测试阶段，将全局特征和局部特征沿通道维度进行拼接表示行人特征，并使用欧氏距离计算行人之间的距离。该方法在Market-1501、DukeMTMC-reID、CUHK03-Labeled和CUHK03-Detected数据集上，mAP分别达到了89.2%、79.7%、80.1%和76.6%，Rank-1分别达到了95.8%、90.4%、82.4%和80.1%。实验结果说明了该算法的识别精度和平均正确率优于当前很多主流算法。
|
| 关 键 词： | ??行人重识别 加权特征融合 深度学习 卷积神经网络 |
| 收稿时间： | 2022-03-08 |
Person Re-identification Method Based on Multi-scale Weighted Feature Fusion |
| |
| Affiliation: | School of Electronic and Information Engineering，Liaoning Technical University，Huludao，Liaoning 125000，China |
| |
| Abstract: | ? ?Aiming at the problem of low person re-identification rate due to the common occlusion and variable pedestrian posture in person re-identification， a person re-identification method based on multi-scale weighted feature fusion（MSWF） was proposed. Firstly， the method used the backbone network ResNeSt-50 to extract downsampling 3 times features， downsampling 4 times features and downsampling 5 times features， and input them into the weighted feature pyramid network and then used a fast normalization fusion method for feature fusion. The introduction of weighting operations in feature fusion allowed the model to learn how to assign weight values to fused features during the training process， so that the features of different scales could be fully utilized to obtain richer person features. Finally， the fused high-level features which had rich semantic information were used as global features， and the fused high-resolution features were used as local features. During the training process， the model was trained by combining Softmax classification loss function， triplet loss function and center loss function. In the testing phase， global features and local features were concatenated along the channel dimension to represent person features， and Euclidean distance was used to calculate distance between preson. A lot of experiments were done on the Market-1501 datasets， DukeMTMC-reID datasets， CUHK03-Labeled datasets and CUHK03-Detected datasets. On the Market-1501 datasets， the mAP and Rank-1 reached 89.2% and 95.8%. On the DukeMTMC-reID datasets， the mAP and Rank-1 reached 79.7% and 90.4%. On the CUHK03-Labeled datasets， the mAP and Rank-1 reached 80.1% and 82.4%. Rank-1 and mAP of CUHK03-Detected datasets reached 80.1% and 76.6%. The experimental results show that the person re-identification rate and mean average precision of this method are better than many current mainstream methods. |
| |
| Keywords: | |
|
| 点击此处可从《信号处理》浏览原始摘要信息 |
|
点击此处可从《信号处理》下载全文 |
