联合归一化模块和多分支特征的行人重识别
Joint normalization module and multi-branch features for person re-identification
任丹萍 1董会升 1何婷婷 1张春华2
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
- 2. 河北工程大学体育与健康学院,河北邯郸 056038
- 折叠
摘要
针对行人重识别技术中存在特征挖掘不充分的问题,提出一种联合归一化模块和多分支特征的行人重识别模型.在主干网络中嵌入注意力机制引导的实例归一化模块,减轻背景等杂波信息的影响.在双级特征融合模块对局部特征进行加权后再聚合形成对行人特征的更细节表达.联合平滑交叉熵损失、三元组损失以及跨分支特征蒸馏损失对网络进行优化.所提模型在Market-1501和DukeMTMC-ReID数据集上首位准确率分别达到了 95.7％和89.2％.实验结果表明,该模型增强了对图像特征的提取.
Abstract
To address the problem of inadequate feature mining in person re-identification techniques,a person re-identification model with a joint normalization module and multi-branch features was proposed.An attention mechanism-guided instance nor-malization module was embedded in the backbone network to mitigate the influence of background and other clutter information.The local features were weighted and then aggregated in a two-level feature fusion module to form a more detailed representation of person features.The network was optimized by jointly smoothing the cross-entropy loss,the triplet loss and the cross-branch feature distillation loss.The proposed model achieves the first accuracy of 95.7％and 89.2％on the Market-1501 and DukeMTMC-ReID datasets respectively.Results of experiments show that the model enhances the extraction of image features.
关键词
归一化/行人重识别/注意力机制/多分支特征/特征提取/特征蒸馏损失/三元组损失Key words
normalization/person re-identification/attention mechanism/multi-branch features/feature extraction/feature dis-tillation loss/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