Robust Degradation Representation via Efficient Diffusion Model for Blind Super-Resolution

Blind super-resolution (SR) is a challenging low-level vision task, dedicated to recovering corrupted details in low-resolution (LR) images with complex unknown degradations. The mainstream blind SR methods mainly adopt the paradigm of capturing the robust degradation representation from the LR images as condition and then perform deep feature reconstruction. However, the manifold degradation factors make it challenging to achieve flexible estimation. In this paper, we propose a residual-guided diffusion degradation representation scheme (Diff-BSR) for blind SR. Specifically, we leverage the powerful generative capability of the diffusion model (DM) to implicitly model the diverse degradations representation, which helps to resist to the disturbance of varied input. Meanwhile, to reduce the expensive computational complexity and training costs, we design a lightweight degradation extractor in the residual domain. It transforms the target residual distribution in a low-dimension feature space. As a result, Diff-BSR requires only about 60 sampling steps and a much smaller scale denoising network. Moreover, we designed the Degradation-Aware Multihead Self-Attention mechanism to effectively fuse the discriminative representations with the intermediate features of the network for robustness enhancement. Extensive experiments on mainstream blind SR benchmarks show that Diff-BSR achieves SOTA or comparable performance compared to existing methods.