Reproduce “LightBSR: Towards Lightweight Blind Super-Resolution via Discriminative Implicit Degradation Representation Learning” on OpenResearch

About this paper

Jiang Yuan, JI Ma, Bo Wang, Guanzhou Ke, Weiming Hu

LightBSR introduces a lightweight blind super-resolution model that enhances image reconstruction by learning highly discriminative implicit degradation representations. This approach enables a simplified adaptation module, achieving state-of-the-art performance with significantly reduced parameters and computational costs across various degradation settings and real-world images.

Key insights

• Prioritizing the discriminability of Implicit Degradation Representations (IDRs) allows for a simpler and more efficient degradation adaptation module, challenging the trend of increasing model complexity for better performance.
• Combining Degradation Reference Priors with a contrastive learning framework effectively enhances the separability of different degradation types in the latent space, which is crucial for robust blind super-resolution.
• A two-stage knowledge distillation approach successfully transfers complex degradation modeling capabilities from a sophisticated teacher to a lightweight student, balancing performance with practical efficiency.

Summary by alphaXiv.