1D kernel distillation network for efficient image super-resolution

Yusong Li

doi:10.1016/j.imavis.2024.105411

Return

1D kernel distillation network for efficient image super-resolution

2025-02-01

PRE

DOI：10.1016/j.imavis.2024.105411

Original

Original request for help

Save

Abstract

En 中文

Recently, there have been significant strides in single-image super-resolution, especially with the integration of transformers. However, the escalating computational demands of large models pose challenges for deployment on edge devices. Therefore, in pursuit of Efficient Image Super-Resolution (EISR), achieving abetter balance between task computational complexity and image fidelity becomes imperative. In this paper, we introduce the 1D kernel distillation network (OKDN). Within this network, we have devised a lightweight 1D Large Kernel (OLK) block, incorporating a more lightweight yet highly effective attention mechanism. This block significantly expands the effective receptive field, enhancing performance while mitigating computational costs. Additionally, we develop a Channel Shift Enhanced Distillation (CSED) block to improve distillation efficiency, allocating more computational resources towards increasing network depth. We utilize methods involving partial channel shifting and global feature supervision (GFS) to further augment the effective receptive field. Furthermore, we introduce learnable Gaussian perturbation convolution (LGPConv) to enhance the model's feature extraction and performance capabilities while upholding low computational complexity. Experimental results demonstrate that our proposed approach achieves superior results with significantly lower computational complexity compared to state-of-the-art models. The code is available at https://github.com/ satvio/OKDN.

Keywords:

Image super-resolution

Large kernel attention

Information distillation

Lightweight