.png){kind=icon}
.png){kind=icon}
Return
Unsupervised image super-resolution recurrent network based on diffusion model
.png){kind=icon}
PRE
AI
DOI：10.1016/j.image.2025.117398.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• This paper uses a cyclic structure to generate paired images from unpaired ones. A color consistency loss function is implemented to ensure the generated image pairs have realistic colors and distributions. • The paper introduces two independent diffusion models to learn both image reconstruction and degradation processes. These models enhance detail generation, resulting in images with richer details. • An end-to-end training approach is used to optimize both branches. This allows the pseudo-paired images to assist the diffusion models in learning the HR image distribution while enabling the diffusion model to improve the recurrent network’s output.
Keywords:
cyclic structure
color consistency loss
diffusion models
image reconstruction
end-to-end training
Journal
S
IF:
2.7
Papers:
2.8K
Citations:
4.2K