联合注意力机制和多分支特征的行人重识别
Person re-identification combined with attention mechanism and multi-branch features
任丹萍 1董会升 1何婷婷1
作者信息
- 1. 河北工程大学信息与电气工程学院,河北邯郸 056038;河北工程大学河北省安防信息感知与处理重点实验室,河北邯郸 056038
- 折叠
摘要
针对行人重识别技术中存在模型识别率低的问题,提出一个联合注意力机制和多分支特征的网络模型.在残差网络中嵌入自注意力机制模块强化图像有效特征的提取,在深度特征挖掘模块,使用全局特征分支、局部关联特征分支以及随机擦除特征分支形成对行人更全面的描述.在优化过程中提出联合余弦交叉熵损失、全样本三元组损失、中心损失以及特征对齐损失对网络使用最小最大策略进行更新.所提方法在Market-1501和DukeMTMC-reID数据集上首位准确率分别达到了 95.8％和 89.8％.
Abstract
To address the problem of low model recognition rate in person re-identification techniques,a network model with joint attention mechanism and multi-branch features was proposed.A self-attentive mechanism module was embedded in the residual network to enhance the extraction of effective features from images,and a global feature branch,a local correlation feature branch and a random erasure feature branch were used in the deep feature mining module to form a more comprehensive descrip-tion of pedestrians.A joint cosine softmax loss,full-sample triplet loss,center loss and feature alignment loss were proposed in the optimization process to update the network using a min-max strategy.The proposed method achieves the first accuracy of 95.8％and 89.8％on the Market-1501 and DukeMTMC-reID datasets respectively.
关键词
行人重识别/深度学习/注意力机制/多分支特征/局部特征/随机擦除/三元组损失Key words
person re-identification/deep learning/attention mechanism/multi-branch feature/local feature/random erasing/triplet loss引用本文复制引用
基金项目
国家重点研发计划基金项目(2018YFF0301004)
国家自然科学基金项目(6210011890)
国家自然科学基金项目(62071071)
河北省自然科学基金项目(F2021402005)
河北省高等学校科学技术研究基金项目(QN2020193)
出版年
2024
NETL
NSTL
万方数据