二值化图像与双流网络在跨模态行人重识别的应用  

APPLICATION OF BINARY IMAGE AND DUAL FLOW NETWORK IN CROSS MODAL RE-ID

作　　者：崔洪刚[1] 曹钢钢 苏荻翔 Cui Honggang;Cao Ganggang;Su Dixiang(School of Computer,Guangdong University of Technology,Guangzhou 510006,Guangdong,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《计算机应用与软件》2025年第2期216-226,共11页Computer Applications and Software

基　　金：广东省东源县优势产业科技创业孵化育成链条项目(2017A010109023)。

摘　　要：在现有的跨模态行人重识别方法中,很少有方法会利用图像中人的姿态信息进行网络的学习。考虑到姿态信息在行人重识别网络学习中的重要性,提出一种融合局部阈值二值化图像特征的端到端的行人重识别方法。该方法使用ResNet50作为骨干网络对三种模态图像进行特征提取和特征融合,使用交叉熵损失和改进的难样本三元组损失进行网络训练。在使用简单网络结构的同时使用姿态信息。实验结果表明,在跨模态行人重识别网络中融合局部阈值二值化图像信息,能提高网络对行人重识别的准确率,显著提升最难样本的挖掘能力。In the existing cross modal Re-ID(Person Re-identification)theories,few theories can use the attitude information in the images to go on the network learning.Considering the importance of attitude in Re-ID network learning,an end-to-end Re-ID theory is proposed which integrates local threshold binary image features.This method used ResNet50 as the backbone network to extract and fusion the features of three modal images.The network was trained by the cross entropy loss and improved hard triplet loss.The attitude information was used while using a simple network structure.The experimental results show that,the fusion of local threshold binary image information in the cross modal Re-ID network can enhance the accuracy of re-identification and improve the ability of finding the most difficult samples.

关 键 词：跨模态行人重识别 卷积神经网络 局部阈值二值化 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]