LKSMN: Large Kernel Spatial Modulation Network for Lightweight Image Super-Resolution

doi:10.1007/s00371-024-03562-5

LKSMN: Large Kernel Spatial Modulation Network for Lightweight Image Super-Resolution

2024-07-25

2

PRE

AI

DOI：10.1007/s00371-024-03562-5

原文链接

原文求助

摘要

En 中文

Although Vits-based networks have achieved stunning results in image super-resolution, their self-attention (SA) modeling in the unidimension greatly limits the reconstruction performance. In addition, the high consumption of resources for SA limits its application scenarios. In this study, we explore the working mechanism of SA and redesign its key structures to retain powerful modeling capabilities while reducing resource consumption. Further, we propose large kernel spatial modulation network (LKSMN); it can leverage the complementary strengths of attention from spatial and channel dimensions to mine a fuller range of potential correlations. Specifically, three effective designs were included in LKSMN. First, we propose multi-scale spatial modulation attention (MSMA) based on convolutional modulation (CM) and large-kernel convolution decomposition (LKCD). Instead of generating feature-relevance scores via queries and keys in the SA, MSMA uses LKCD to act directly on the input features to produce convolutional features that imitate relevance scores matrix. This process reduces the computational and storage overhead of the SA while retaining its ability to robustly model long-range dependent correlations. Second, we introduce multi-dconv head transposed attention (MDTA) as an attention modeling scheme in the channel dimension, which complements the advantages of our MSMA to model pixel interactions in both dimensions simultaneously. Final, we propose a multi-level feature aggregation module (MLFA) for aggregating the feature information extracted from different depth modules located in the network, to avoid the problem of shallow feature information disappearance. Extensive experiments demonstrate that our proposed method can achieve competitive results with a small network scale (e.g., 26.33dB@Urban100 x\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} 4 with only 253K parameters). The code is available at https://figshare.com/articles/software/LKSMN_Large_Kernel_Spatial_Modulation_Network_for_Lightweight_Image_Super-Resolution/25603893

Keyword:

Image super-resolution

Self-attention

Convolutional modulation

Large-kernel convolution decomposition

LKSMN: Large Kernel Spatial Modulation Network for Lightweight Image Super-Resolution

摘要

期刊

机构

引用论文