【图像超分】论文精读:SeD: Semantic-Aware Discriminator for Image Super-Resolution
第一次来请先看这篇文章:【超分辨率(Super-Resolution)】关于【超分辨率重建】专栏的相关说明,包含专栏简介、专栏亮点、适配人群、相关说明、阅读顺序、超分理解、实现流程、研究方向、论文代码数据集汇总等)
文章目录
- 前言
- Abstract
- 1. Introduction
- 2. Related Works
- 2.1. Single Image Super-resolution
- 2.2. GAN-based Image Super-resolution
- 2.3. Pretrained Vision Models (PVMs)
- 3. Method
- 3.1. Preliminary
- 3.2. Overall Framework
- 3.3. Semantic-aware Discriminator
- 3.3.1 Semantic Excavation
- 3.3.2 Semantic-aware Fusion Block
- 3.4. Extension to Various Discriminators
- 4. Experiments
- 4.1. Experiment setups
- 4.2. Results on classical image SR
- 4.3. Results on real-world image SR
- 4.4. Ablation studies for SeD
- 4.5. T-SNE visualization of discriminator features
- 5. Conclusion
- Appendix
- 6. Evaluation on large-scale benchmarks
- 7. Implementation detials of image-wise SeD
- 8. Implementation details of different fusion strategies
- 9. Ablation studies with Semantic-aware Generator
- 10. More visualization results
前言
论文题目:SeD: Semantic-Aware Discriminator for Image Super-Resolution —— SeD:图像超分辨率语义感知鉴别器
论文地址:SeD: Semantic-Aware Discriminator for Image Super-Resolution
论文源码:https://github.com/lbc12345/SeD
CVPR 2024!
Abstract
生成对抗网络 (GAN) 已被广泛用于恢复图像超分辨率 (SR) 任务中生动的纹理。特别是,一个鉴别器被用来使SR网络能够以对抗训练的方式学习真实世界高质量图像的分布。然而,分布学习过于粗粒度,容易受到虚拟纹理的影响,并导致违反直觉的生成结果。为了缓解这种情况,我们提出了简单有效的语义感知鉴别器(表示为 SeD),它鼓励 SR 网络通过引入图像的语义作为条件来学习细粒度分布。具体来说,