基于师生网络协同的动态阈值半监督学习图像分类研究

打开文本图片集

中图分类号：TP273；TP242 文献标识码：A 文章编号：2096-4706（2025）20-0055-06

Abstract： The application of semi-supervised learning in image processing receives wide atention. Among them， the method basedonconsistencyregularizationand pseudo-label improves the performanceof the model by effectivelyusing a largeamountofunlabeleddata.However，mostoftheexisting methods use high fixedthresholds to generatepseudo-labels forunlabeleddatawhichmakesitdiffculttofectivelyutilizenlabeleddata.Inadition，theteachernetworkmaymislead the student network due to performance fluctuations. Aiming at the above problems，this paper proposes a dynamic threshold semi-supervised learning image clasificationmethod MTDT.This methodonlysupervises studentnetworktrainingbyteacher network whenteachernetworkperformanceisbettr thanstudentnetwork.Atthesametime，accordingtotheleaingstateof themodel， theconfidencethresholdisdyamicalldjustedinandaptivemaner.TeexperimentalsultsonCFA-1，ST10 and SVHN public datasets show that the proposed method has higher classification accuracy.

Keywords： adaptive threshold; pseudo-label; semi-supervised; consistency regularizatior

0 引言

深度学习之所以能表现出优异性能，很大程度上依赖于充足的标记数据对模型进行监督训练。（剩余9958字）