基于UPS策略自我训练的半监督语义分割  

Semi-supervised Semantic Segmentation Based on UPS Strategy Self Training

作　　者：李雨杭 朱小东 杨高明[1] LI Yuhang;ZHU Xiaodong;YANG Gaoming(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《现代信息科技》2024年第2期1-4,共4页Modern Information Technology

基　　金：安徽高校自然科学研究项目(KJ2017A084);安徽省自然科学基金面上项目(1808085MF179)。

摘　　要：为提高半监督语义分割的效果,文章提出一种损失归一化技术结合UPS策略的半监督语义分割网络SPNS。利用损失归一化技术缓解标准损失函数的自我训练不稳定问题;UPS策略是结合不确定性估计和消极学习的技术,通过计算输出值的不确定性作为另一种阈值,用以挑选可靠的伪标签,最后利用生成的伪标签和标记数据完成半监督语义分割任务。SPNS方法在PASCAL·VOC数据集上相对于只使用标记数据训练有着+2.06的效果提升,与其他方法相比也有一定提升。To improve the effectiveness of semi-supervised semantic segmentation,this paper proposes a semi-supervised semantic segmentation network SPNS that combines loss normalization technology with UPS strategy.Using loss normalization techniques to alleviate the instability of self training in standard loss functions;the UPS strategy is a technique that combines uncertainty estimation and passive learning,by calculating the incompleteness of the output value as another threshold,reliable pseudo labels are selected,and finally the semi-supervised semantic segmentation task is completed using the generated pseudo labels and labeled data.The SPNS method has+2.06 improvement compared to training with only labeled data on the PASCAL•VOC dataset,and also has some improvement compared to other methods.

关 键 词：半监督 语义分割 自我训练 UPS 消极学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]