文摘
|
行人重识别旨在跨监控设备下检索出特定的行人目标.为捕捉行人图像的多粒度特征进而提高识别精度,基于OSNet基准网络提出一种多粒度特征融合网络(Multi-granularity Feature Fusion Network for Person Re-Identification, MFN)进行端对端的学习. MFN由全局分支、特征擦除分支和局部分支组成,其中特征擦除分支由双通道注意力擦除模型构成,此模型包含通道注意力擦除模块(Channel Attention-based Dropout Moudle, CDM)和空间注意力擦除模块(Spatial Attention-based Dropout Moudle, SDM). CDM对通道的注意力强度排序并擦除低注意力通道,SDM在空间维度上以一定概率擦除最具有判别力的特征,两者通过并联方式相互作用,提高模型的识别能力.全局分支采用特征金字塔结构提取多尺度特征,局部分支将特征均匀切块后级联成一个单一特征,提取关键局部信息.大量实验结果表明了本文方法的有效性,在Market1501、DukeMTMC-reID和CUHK03-Labeled(Detected)数据集上,mAP/Rank-1分别达到了90.1%/95.8%、81.8%/91.4%和80.7%/82.3%(78.7%/81.6%),大幅优于其他现有方法. |
其他语种文摘
|
For the purpose of capturing the multi-granularity features and improving the recognition accuracy, a multi-granularity feature fusion network for person re-identification (MFN) is proposed based on the omist-scale network (OSNet). The MFN network is composed of a global branch, a feature dropout branch and a local branch. The feature dropout branch consists of a dual-channel attention dropout model, which includes a channel attention-based dropout moudle (CDM) and a Spatial attention-based dropout moudle (SDM). CDM sorts the attention intensity and dropouts low attention channels, and SDM dropouts the most discriminative features with a certain probability in the spatial dimension. The global branch uses the feature pyramid structure to extract multi-scale features, and the local branch employs a uniform partition strategy to produce local features which are cascaded into a single one for key local information extraction. Experiments on the large scale datasets show the effectiveness of MFN. On the Market1501, DukeMTMC-reID and CUHK03 -Labeled (Detected) datasets, mAP/Rank-1 of MFN reaches 90.1%/95.8%, 81.8%/91.4% and 80.7%/82.3% (78.7%/81.6%), which is superior to other existing methods. |
来源
|
电子学报
,2021,49(8):1541-1550 【核心库】
|
DOI
|
10.12263/DZXB.20200974
|
关键词
|
行人重识别
;
多分支CNN网络
;
金字塔结构
;
特征擦除
|
地址
|
江南大学, 轻工过程先进控制教育部重点实验室, 江苏, 无锡, 214122
|
语种
|
中文 |
文献类型
|
研究性论文 |
ISSN
|
0372-2112 |
学科
|
自动化技术、计算机技术 |
基金
|
国家自然科学基金
|
文献收藏号
|
CSCD:7055897
|