基于对抗学习的开集域自适应分类
Open-set domain adaptation classification by adversarial learning
  
DOI:
中文关键词:  开集域自适应; 对抗学习;图像分类;迁移学习
英文关键词:open-set domain adaptation; adversarial learning; image classification;transfer learning
基金项目:
作者单位
张庆亮 南京邮电大学 自动化学院、人工智能学院,江苏 南京 210023 
朱松豪 南京邮电大学 自动化学院、人工智能学院,江苏 南京 210023 
摘要点击次数: 2638
全文下载次数: 2021
中文摘要:
      通过域自适应方法,利用标记的源域样本实现未标记目标域样本的识别,目前逐渐成为机器视觉领域一个新的研究热点。与传统的闭集域自适应问题不同,开集域自适应中的目标域包含源域中未出现的类别,因此该问题的研究更为贴近实际场景,但同时也增加了识别难度。文中提出了一种基于反向传播开集域自适应的方法,该方法利用对抗思想实现了域不变特征的提取,同时通过设置阈值实现目标域中未知类别样本的消除。在两个公共域自适应数据集上的大量实验结果表明,与以前的方法相比,文中提出的奇异值平衡提高了模型的识别率,在一定程度上有效解决了开集域自适应的问题。
英文摘要:
      A domain adaptation method is used for labeled source domain samples to identify unlabeled target domain samples, it has become a new research hotspot in the field of machine vision. Different from the traditional closed-set domain adaptation problem, the target domain in the open-set domain adaptation contains categories that do not appear in the source domain. Therefore, the problem is closer to the actual scene, but it also increases the difficulty of recognition. A method based on the open set domain adaptation by back-propagation is proposed. The method uses adversarial ideas to extract domain invariant features and eliminate the unknown category samples in the target domain by setting a threshold. Experimental results on two public domain adaptation datasets show that compared with the previous methods, the singular value balance proposed in this paper improves the recognition rate of the model. Thus, the method solves the problem of the open set domain adaptation to a certain extent.
查看全文  查看/发表评论  下载PDF阅读器

你是第3766563访问者
版权所有《南京邮电大学学报(自然科学版)》编辑部
Tel:86-25-85866913 E-mail:xb@njupt.edu.cn
技术支持:本系统由北京勤云科技发展有限公司设计