基于课程学习的域自适应伪标签样本选择算法  

Curriculum Learning Based Pseudo-Label Sample Selection Algorithm for Domain Adaptation

作　　者：温颖涵 曹江中[1] WEN Ying-han;CAO Jiang-zhong(School of Information Engineering,Guangdong University of Technology,Guangzhou Guangdong 510006,China)

机构地区：[1]广东工业大学信息工程学院,广东广州510006

出　　处：《计算机仿真》2024年第10期204-209,共6页Computer Simulation

基　　金：国家自然科学基金-广东联合基金项目(U1701266)。

摘　　要：针对域自适应网络在不同学习阶段对不同样本学习能力存在差异性问题,提出一种基于课程学习的域自适应伪标签样本选择算法。算法利用置信值将目标域数据集中的易学习样本转入源数据集,解决了模型前期学习能力不足可能造成的错误累积问题;同时通过模型的学习状态更新各类样本的伪标签阈值,得到无监督条件下难易学习样本类估计,输出可靠的置信值并减少噪声伪标签对模型学习的影响。提出的算法在三个主流无监督域自适应目标检测数据集上进行了仿真,仿真结果表明,上述算法相比最近提出的非对抗或增强的无监督域自适应方法在精确度上得到了有效提升。A domain adaptive pseudo label sample selection algorithm based on curriculum learning is proposed to address the issue of differences in learning ability of domain adaptive networks for different samples at different learning stages.The algorithm transfers the easy-to-learn samples in the target domain dataset into the source dataset through the confidence value,and solves the problem of error accumulation caused by the insufficient learning ability of the model in the early stage;At the same time,the pseudo-label threshold of all kinds of samples is updated by the learning state of the model,and the estimation of sample class that is difficult to learn under unsupervised conditions is obtained,and the reliable confidence value is output and reduced the influence of noise false labels on model learning.The proposed algorithm is verified on three mainstream unsupervised domain adaptive target detection data sets.The experimental results show that the accuracy of the proposed algorithm is effectively improved compared with the recently proposed non-confrontational or enhanced unsupervised domain adaptation method.

关 键 词：域自适应 课程学习 伪标签 目标检测 置信度估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]