结合域间一致性与抗噪学习的跨域行人重识别方法  

Cross-domain Person Re-identification Method Combining Inter-domain Consistency and Anti-noise Learning

作　　者：张栩嘉 白淼源 郭继峰[1] ZHANG Xujia;BAI Miaoyuan;GUO Jifeng(College of Information and Computer Engineering,Northeast Forestry University,Harbin 150000,China)

机构地区：[1]东北林业大学信息与计算机工程学院,哈尔滨150000

出　　处：《小型微型计算机系统》2024年第6期1444-1450,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61300098)资助;黑龙江省自然科学基金项目(LH2019C003)资助;中央高校基本科研业务费专项资金项目(2572019BH03)资助。

摘　　要：针对跨域行人重识别任务中域间差距和噪声标签导致的模型识别准确率低、稳定性差的问题,本文提出一种结合域间一致性与抗噪学习的跨域行人重识别方法(SZTR).该方法将生成对抗网络应用到预训练中,并且设计一个预训练中的一致性正则化损失,用来保持行人图片风格迁移前后预测概率的一致性,以减小域间差距.在目标域的训练中引入更稳定的动量对比模型来提取样本特征,加入鲁棒性更强的自适应权重损失微调模型,以减小聚类过程中噪声标签带来的负面影响.实验结果及分析表明,所提方法有助于提升模型在跨域行人重识别任务上的准确度,取得了更稳定的识别效果.Aiming at the problem of low recognition accuracy and poor stability of model caused by inter-domain gap and noise label in cross-domain person re-identification task,this paper proposes a cross-domain person re-identification method combining inter-domain consistency and anti-noise learning(SZTR).In this method,the generative adversarial network is applied to the pre-training,and a regularization loss in the pre-training is designed to maintain the consistency of the prediction probability before and after the style transfer of pedestrian images,so as to reduce the inter-domain gap.In the training of the target domain,a more stable momentum comparison model is introduced to extract sample features,and a more robust adaptive weight loss fine-tuning model is added to reduce the negative impact of noise labels in clustering process.The experimental results and analysis show that the proposed method helps to improve the accuracy of the model in cross-domain person re-identification tasks,and achieves a more stable recognition effect.

关 键 词：跨域 行人重识别 生成对抗网络 一致性 噪声标签 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]