摘要
随着深度学习在目标检测中的应用越来越广泛,其对标记数据的过度依赖问题日益凸显。为了缓解目标检测任务对高成本标记数据的依赖,半监督学习算法备受关注且飞速发展。如何有效挖掘无标数据中蕴含的信息,并获取高质量的伪标签,是当前半监督目标检测算法面临的问题。对此,提出双分支的教师-学生(Teacher-Student,TS)网络结构缓解传统单分支教师-学生网络结构中的耦合问题,有效挖掘、利用无标数据。基于所提出的双分支网络结构设计一种伪标签生成机制,以获取高质量的伪标签用于模型训练,从而提升模型性能。在PASCALVOC数据集和MSCOCO数据集上分别进行多组实验,结果表明提出的方法能有效提升模型性能。
With the wide application of deep learning in target detection,the problem of over-dependence on labeled data becomes more and more prominent.In order to alleviate the dependence of target detection task on high-cost labeled data,semi-supervised learning algorithm has been paid much attention and developed rapidly.How to effectively mine the information contained in unmarked data and how to obtain high quality false labels are the problems faced by the current semi-supervised object detection algorithms.In this paper,a two-branch Teacher-Student(TS)network structure is proposed to alleviate the coupling problem in the traditional single-branch teacher-student network structure and effectively mine and utilize unmarked data.Based on the proposed two-branch network architecture,a pseudo-label generation mechanism is designed to obtain high-quality pseudo-tags for model training and improve model performance.Several experiments were carried out on PASCALVOC and MSCOCO data sets respectively,and the results show that the proposed method can effectively improve the model performance.
作者
叶维裕
陈景
YE Weiyu;CHEN Jing(Chongzuo Preschool Education College,Chongzuo 532200,China)
出处
《智能物联技术》
2025年第1期96-101,共6页
Technology of Io T& AI
基金
2023年崇左幼儿师范高等专科学校校级科研课题(2023XB03)。
关键词
目标检测
半监督学习
伪标签
数据集
object detection
semi supervised learning
false labels
data set
