| 基于元学习的无监督风格迁移算法研究 |
| |
| 引用本文： | 李鑫然.基于元学习的无监督风格迁移算法研究[J].移动信息．新网络,2023,45(6):213-215. |
| |
| 作者姓名： | 李鑫然 |
| |
| 作者单位： | 江南大学人工智能与计算机学院 江苏 无锡 214122 |
| |
| 摘 要： | 最近,在生成式对抗网络和足够的非配对训练数据下,无监督领域风格迁移取得了较高的性能。然而，现有的领域迁移框架大多基于庞大的训练数据集，且只能根据训练图像进行特定类别的风格迁移，忽略了其中的学习经验被，使获得的模型不能适应新的领域。文中对传统的非配对循环生成对抗网络Cycle-GAN进行了改进，并使用元学习方法训练了无监督领域的风格迁移问题。另外，文中提出的模型在7个不同的双域迁移任务上证明了其有效性，当对每个新领域进行小样本训练时，该算法均优于传统的风格迁移算法。
|
| 关 键 词： | 风格迁移 生成对抗网络 元学习 |
| 收稿时间： | 2023/4/10 0:00:00 |
Research on Unsupervised Style Transfer Algorithm Based on Meta-learning |
| |
| LI Xinran.Research on Unsupervised Style Transfer Algorithm Based on Meta-learning[J].Mobile Information,2023,45(6):213-215. |
| |
| Authors: | LI Xinran |
| |
| Affiliation: | School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi,Jiangsu 214122 ,China |
| |
| Abstract: | Recently, unsupervised domain style transfer has achieved high performance under generative adversarial networks and sufficient unpaired training data. However, most of the existing domain transfer frameworks are based on huge training datasets, and can only perform specific classes of style transfer based on training images, where the learning experience is ignored, and make the obtained models cannot be adapted to new domains. In this paper, the traditional unpaired cyclic generative adversarial network Cycle-GAN is improved, and the style transfer problem in unsupervised domains is trained using meta-learning methods. In addition, the model proposed in this paper proves its effectiveness on 7 different dual-domain transfer tasks, when training on small samples for each new domain, the algorithm outperforms traditional style transfer algorithms. |
| |
| Keywords: | |
|
| 点击此处可从《移动信息．新网络》浏览原始摘要信息 |
|
点击此处可从《移动信息．新网络》下载全文 |
|
