Large Kernel Distillation Network for Efficient Single Image Super-Resolution

TL;DR

LKDN tackles the efficiency gap in single-image super-resolution by integrating a large-kernel design with distillation-style blocks, replacing heavy attention modules with a lightweight Large Kernel Attention (LKA), and employing re-parameterization alongside the Adan optimizer to boost training and inference performance. The method achieves state-of-the-art results among efficient SR methods while keeping model size and computations low, with LKDN-S offering competitive NTIRE-ready efficiency. Ablation studies validate the benefits of LKA, the proposed re-parameterization, and the Adan optimizer for faster convergence and improved accuracy. The approach has practical impact for real-time and mobile SR tasks where computational budgets are constrained.

Abstract

Efficient and lightweight single-image super-resolution (SISR) has achieved remarkable performance in recent years. One effective approach is the use of large kernel designs, which have been shown to improve the performance of SISR models while reducing their computational requirements. However, current state-of-the-art (SOTA) models still face problems such as high computational costs. To address these issues, we propose the Large Kernel Distillation Network (LKDN) in this paper. Our approach simplifies the model structure and introduces more efficient attention modules to reduce computational costs while also improving performance. Specifically, we employ the reparameterization technique to enhance model performance without adding extra cost. We also introduce a new optimizer from other tasks to SISR, which improves training speed and performance. Our experimental results demonstrate that LKDN outperforms existing lightweight SR methods and achieves SOTA performance.

Figures (6)

• Figure 1: Comparison of model performance and complexity on Urban100 huang2015single with SR($\times4$).
• Figure 2: The architecture of large kernel distillation network (LKDN).
• Figure 3: The details of each component. (a) LKDB: Large Kernel Distillation Block; (b) BSConv: Blueprint Separable Convolution; (c) LKA: Large Kernel Attention; (d) RBSB: Re-parameterized Blueprint Shallow Block.
• Figure 4: Non-reparameteried components. (a) BSRB: Blueprint Shallow Residual Block; (b) BSB: Blueprint Shallow Block.
• Figure 5: Convergence comparison between Adan xie2022adan and Adam kingma2014adam optimizers, using Ubran100 huang2015single SR($\times4$).