| 姚明海,赵彦涛,项圣,顾勤龙.基于改进自训练的半监督学习算法研究[J].高技术通讯(中文),2025,35(3):330~338 |
| 基于改进自训练的半监督学习算法研究 |
| Research on semi-supervised learning algorithm based on improved self-training |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 03. 010 |
| 中文关键词: 半监督学习； 自训练； 伪标签筛选； 稳定教师网络 |
| 英文关键词: semi-supervised learning, self-training, pseudo-label selective strategy, stable teacher network |
| 基金项目: |
| 作者 | 单位 | | 姚明海 | （浙江工业大学信息工程学院杭州 310023） | | 赵彦涛 | | | 项圣 | | | 顾勤龙 | |
|
| 摘要点击次数: 396 |
| 全文下载次数: 264 |
| 中文摘要: |
| 本文针对半监督方法中存在的标注及未标注数据利用率不高、训练时间长等问题，提出了一种基于改进自训练的半监督语义分割方法。首先，针对网络训练时间成本较高的问题，提出一个稳定教师网络选取策略，可以根据网络训练中稳定性自动选择网络训练的停止时间，有效提高训练效率；其次，对未标注数据采取多重随机数据增强方法，扩充数据集，有效防止训练过程中过拟合现象；最后，对网络生成的伪标签采取筛选策略，为重训练阶段筛选高质量伪标签，提高模型的分割效果。实验对比结果表明，使用本方法在Pascal VOC 2012 SBD数据集上进行验证时，在比例分别为1/16、1/8、1/4的有标注数据集下相应的平均交并比(mean intersection over union，mIOU)达到了72.7%、74.3%、75.4%，而且在保证分割精度的情况下，网络训练效率提升了近30.0%。 |
| 英文摘要: |
| In this paper, an improved self-training-based semi-supervised semantic segmentation method is proposed to address challenges in semi-supervised approaches, including low utilization of labeled and unlabeled data and extended training times. Firstly, to tackle the issue of high time costs in network training, a stable teacher network selection strategy is introduced. This strategy automatically determines the stopping time for network training based on its stability during training, effectively enhancing training efficiency. Secondly, multiple random data augmentation techniques are applied to unlabeled data to expand the dataset and prevent overfitting during training. Finally, a filtering strategy is implemented for pseudo-labels generated by the network, selecting high-quality pseudo-labels during the retraining phase to improve model segmentation performance. Comparative experimental results demonstrate that employing the proposed method for validation on the Pascal VOC 2012 SBD dataset, with labeled data proportions of 1/16, 1/8, and 1/4, yields corresponding mean intersection over union (mIOU) values of 72.7%, 74.3%, and 75.4%. Notably, while ensuring segmentation accuracy, the efficiency of network training is boosted by almost 30.0%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
