4.6 Article

InvolutionGAN: lightweight GAN with involution for unsupervised image-to-image translation

期刊

NEURAL COMPUTING & APPLICATIONS
卷 35, 期 22, 页码 16593-16605

出版社

SPRINGER LONDON LTD
DOI: 10.1007/s00521-023-08530-z

关键词

Unsupervised image-to-image translation; Image style transfer; Generative adversarial networks; Deep learning

向作者/读者索取更多资源

The unsupervised image-to-image translation aims to learn a mapping that translates images from one domain to another. Current GAN models require expensive operations and suffer from high computational costs. To address this, we propose using involution, a lightweight operator, to enhance the GAN structure and introduce a novel loss term to evaluate perceptual similarity distance.
The unsupervised image-to-image translation aims to learn a mapping that translates images from one domain to the target domain. Current state-of-the-art generative adversarial network (GAN) models utilize time and space-costly operators to produce impressive translated images. However, further research and model deployment are under restrictions due to the high computational costs of the models. In order to resolve the problem, we enhance the GAN structure by employing a lightweight operator named involution that facilitates extracting both local features and long-range dependencies across channels. Besides, we also notice that previous works attach less importance to feature-level reconstruction discrepancy between original and reconstructed images. Nevertheless, such information is crucial in improving the quality of the synthesized images. Thus, we develop a novel loss term that evaluates the learned perceptual similarity distance to regulate the training process. The qualitative and quantitative experiment results on several prevailing benchmarks demonstrate that our model, dubbed InvolutionGAN, could produce competitive image results while saving computational costs up to 91.9%. In addition, extensive ablation studies are conducted to search for the best model structure and verify that each component we introduced is effective.

作者

我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。

评论

主要评分

4.6
评分不足

次要评分

新颖性
-
重要性
-
科学严谨性
-
评价这篇论文

推荐

暂无数据
暂无数据