Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning

Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning

5 Mar 2024 | Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, ZhenSong Zhang, Youliang Yan, Lei Zhu
The paper "Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning" introduces a novel training framework called "Low-Res Leads the Way" (LWay) to enhance the adaptability of super-resolution (SR) models to real-world images. LWay combines supervised pre-training with self-supervised learning, utilizing a low-resolution (LR) reconstruction network to extract degradation embeddings from LR images. These embeddings are then merged with super-resolved outputs for LR reconstruction. The approach leverages unseen LR images for self-supervised learning, guiding the model to adapt its modeling space to the target domain, facilitating fine-tuning of SR models without requiring paired high-resolution (HR) images. The integration of Discrete Wavelet Transform (DWT) further refines the focus on high-frequency details. Extensive evaluations show that LWay significantly improves the generalization and detail restoration capabilities of SR models on unseen real-world datasets, outperforming existing methods. The training regime is universally compatible, requiring no network architecture modifications, making it a practical solution for real-world SR applications. The paper discusses the limitations and comparisons with other methods, emphasizing the robustness and effectiveness of LWay in handling complex and variable degradation patterns.The paper "Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning" introduces a novel training framework called "Low-Res Leads the Way" (LWay) to enhance the adaptability of super-resolution (SR) models to real-world images. LWay combines supervised pre-training with self-supervised learning, utilizing a low-resolution (LR) reconstruction network to extract degradation embeddings from LR images. These embeddings are then merged with super-resolved outputs for LR reconstruction. The approach leverages unseen LR images for self-supervised learning, guiding the model to adapt its modeling space to the target domain, facilitating fine-tuning of SR models without requiring paired high-resolution (HR) images. The integration of Discrete Wavelet Transform (DWT) further refines the focus on high-frequency details. Extensive evaluations show that LWay significantly improves the generalization and detail restoration capabilities of SR models on unseen real-world datasets, outperforming existing methods. The training regime is universally compatible, requiring no network architecture modifications, making it a practical solution for real-world SR applications. The paper discusses the limitations and comparisons with other methods, emphasizing the robustness and effectiveness of LWay in handling complex and variable degradation patterns.
Reach us at info@study.space