Lightweight Image Super-Resolution via Progressive Pruning and Adaptive Distillation

초록

Recent advances in image super-resolution (SR) have achieved remarkable performance through the adoption of transformer-based and diffusion-based models. However, their high computational complexity and memory requirements pose significant challenges for real-world deployment. To address these issues, we propose a two-stage lightweight training framework that integrates model pruning and knowledge distillation. In the first stage, we perform efficient channel reduction using progressive local filter pruning, which balances structural preservation with computational cost reduction. In the second stage, we introduce mask-based adaptive response distillation to overcome the limitations of conventional L1-based response distillation, enabling more effective knowledge transfer focused on informative regions. Experimental results on various benchmark SR datasets demonstrate that the proposed method achieves superior performance and efficiency compared to existing lightweight SR approaches. © 2025 IEEE.

키워드