WDFSR: Normalizing Flow based on Wavelet-Domain for Super-Resolution

Song, Chao; Li, Shaobang; Li, Frederick W. B.; Yang, Bailin

Abstract

We propose a Normalizing flow based on the wavelet framework for super-resolution called WDFSR. It learns the conditional distribution mapping between low-resolution images in the RGB domain and high-resolution images in the wavelet domain to generate high-resolution images of different styles simultaneously. To address the issue of some flow-based models being sensitive to datasets, resulting in training fluctuations that reduce the mapping ability of the model and weaken generalization, we designed a method that combines a T-distribution and QR decomposition layer. This method alleviates the problem while maintaining the model's ability to map different distributions and produce higher quality images. Good contextual conditional features can promote model training and enhance distribution mapping capabilities for conditional distribution mapping. Therefore, we propose a Refinement layer combined with an attention mechanism to refine and fuse the extracted condition features for improving image quality. Extensive experiments on many super-resolution datasets show that WDFSR outperforms most general CNN models and flow-based models in terms of PSNR and perception quality. We also demonstrate that our framework works well for other low-level vision tasks, such as low-light enhancement. The pre-trained models and source code with guidance for reference are available at https://github.com/Lisbegin/WDFSR.

Citation

Song, C., Li, S., Li, F. W. B., & Yang, B. (in press). WDFSR: Normalizing Flow based on Wavelet-Domain for Super-Resolution. Computational Visual Media,

• Journal Article Type: Article
• Acceptance Date: Aug 22, 2023
• Deposit Date: Sep 12, 2023
• Journal: Computational Visual Media
• Print ISSN: 2096-0433
• Electronic ISSN: 2096-0662
• Publisher: SpringerOpen
• Peer Reviewed: Peer Reviewed
• Public URL: https://durham-repository.worktribe.com/output/1735734
• Publisher URL: https://www.springer.com/journal/41095