Reproduce “Large Kernel Distillation Network for Efficient Single Image Super-Resolution” on OpenResearch

About this paper

Chengxing Xie, Xiaoming Zhang, Linze Li, Haiteng Meng, Tianlin Zhang, Tianrui Li, Xiaole Zhao

Researchers at Southwest Jiaotong University and Chinese Academy of Science developed a Large Kernel Distillation Network (LKDN) for single image super-resolution that combines efficient large kernel attention and re-parameterization techniques to achieve state-of-the-art reconstruction quality. This approach reduces model complexity and computational cost, and its integration of the Adan optimizer leads to faster training and improved performance compared to prior lightweight models like BSRN.

Key insights

• Large kernel designs, when efficiently decomposed, can effectively capture broad contextual information and improve receptive fields without high computational costs.
• Re-parameterization techniques offer a practical way to boost model performance during training while maintaining or even reducing complexity for inference.
• Advanced optimizers like Adan can significantly accelerate training convergence and lead to better final model performance in low-level vision tasks like SISR.

Summary by alphaXiv.