Scholarlink

<p>Image-to-image (I2I) translation is a practical task in computer vision that aims to map one image domain to another. In recent years, methods based on generative adversarial networks (GANs) have emerged as a primary approach for this task. While many methods have been proposed for paired data, fewer solutions exist for the unpaired scenario, with CycleGAN being one of the key methods utilizing cycle consistency loss. However, CycleGAN and its extensions rely on two separate generative networks, increasing the number of parameters and complicating the model by performing domain conversions independently. In this paper, we propose a novel approach that addresses these limitations by utilizing parameter sharing across generators and introducing a new loss function named <i>domain alignment</i> (DA) to enhance the performance by bringing the distribution of the domains close together in feature space. Our method effectively reduces the number of parameters, accelerates training time, and achieves superior or comparable results to CycleGAN. Extensive experiments conducted across five I2I tasks—including object transfiguration, season transfer, photo enhancement, day/night conversion, and artistic style transfer—demonstrate that our model outperforms CycleGAN in most domains, particularly under strong appearance shifts such as lighting and texture changes. Quantitative results using Fréchet Inception distance (FID) and kernel Inception distance (KID), along with qualitative analysis and user studies, further validate the visual quality and semantic consistency of our outputs. Additionally, this technique has the potential to be applied to other cycle-based models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cyclegan++: enhancing performance through parameter sharing and domain alignment

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Mohammad Mahmoudabadi,
  • Mansoor Rezghi

摘要

Image-to-image (I2I) translation is a practical task in computer vision that aims to map one image domain to another. In recent years, methods based on generative adversarial networks (GANs) have emerged as a primary approach for this task. While many methods have been proposed for paired data, fewer solutions exist for the unpaired scenario, with CycleGAN being one of the key methods utilizing cycle consistency loss. However, CycleGAN and its extensions rely on two separate generative networks, increasing the number of parameters and complicating the model by performing domain conversions independently. In this paper, we propose a novel approach that addresses these limitations by utilizing parameter sharing across generators and introducing a new loss function named domain alignment (DA) to enhance the performance by bringing the distribution of the domains close together in feature space. Our method effectively reduces the number of parameters, accelerates training time, and achieves superior or comparable results to CycleGAN. Extensive experiments conducted across five I2I tasks—including object transfiguration, season transfer, photo enhancement, day/night conversion, and artistic style transfer—demonstrate that our model outperforms CycleGAN in most domains, particularly under strong appearance shifts such as lighting and texture changes. Quantitative results using Fréchet Inception distance (FID) and kernel Inception distance (KID), along with qualitative analysis and user studies, further validate the visual quality and semantic consistency of our outputs. Additionally, this technique has the potential to be applied to other cycle-based models.