Expanded SPAN for Efficient Super-Resolution

This work proposes ESPAN, an efficient super-resolution (SR) network that extracts robust representations with constrained parameters by incorporating innovations from three perspectives: self-distillation and progressive learning (SDPL), general re-parameterization (GRep), and frequency-aware loss. In detail, SDPL shares partial blocks between the student and teacher models and progressively removes the tail convolutions of the student model, which contributes to a stable training process and reasonable convergence. Regarding GRep, we provide a more general schema of re-parameterization with interpretable theoretical derivation to achieve more flexible expansion of re-parameterization complexity. The frequency-aware loss utilizes the discrete cosine transform and a high-pass filter, enforcing the model to focus more on important high-frequency areas. The experimental results demonstrate the effectiveness of the proposed strategies. Overall, ESPAN exhibits better generality and robustness than previous top-ranking solutions in the NTIRE ESR challenge (e.g., 0.33 dB higher than SPAN on Manga109) while maintaining inference and restoration performance.