Scholarlink

<p>In the current research on super-resolution reconstruction of digital rock images, most networks only utilize attention mechanisms in a single dimension, neglecting the interaction between spatial and channel dimensions, and underutilizing features extracted at different depths. Additionally, employing both spatial and channel attention simultaneously to capture richer latent correlations results in redundant computations, thereby increasing the scale of the network and raising computational costs. To address these challenges, we propose a digital rock image super-resolution reconstruction network incorporating a Lightweight Two-Dimensional Multi-Scale Large Kernel Attention Building Block (LDMB). The LDMB consists of two Metaformer-Style Convolution Building Blocks (MSCB), along with two spatial attention units and channel attention units. The MSCB consists of a Multi-Scale Large Kernel Attention Unit (MLKA), and an innovative Gated Depthwise Convolutional Feedforward Neural Network (GCFN). Notably, we design a combined use of spatial and channel attention by innovatively proposing Enhanced Dense Convolutional Matrix Spatial Attention (EDSA) and Optimized Blueprint Contrast-Aware Channel Attention (OBCA), which integrates channel information from a global perspective, effectively enhancing the model’s ability to capture subtle textures and complex pore structures in digital rock images. Finally, to improve the contrast of the reconstructed rock images, we employ threshold segmentation techniques to process the images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight two dimensional multi-scale large kernel attention network for super-resolution of digital rock

  • 期刊论文

  • 发表日期：

    2025年2月13日

查看全文
  • Yubo Zhang,
  • Junhao Bi,
  • Chao Han,
  • Lei Xu,
  • Haibin Xiang,
  • Haihua Kong,
  • Juanjuan Geng,
  • Wanying Zhao

摘要

In the current research on super-resolution reconstruction of digital rock images, most networks only utilize attention mechanisms in a single dimension, neglecting the interaction between spatial and channel dimensions, and underutilizing features extracted at different depths. Additionally, employing both spatial and channel attention simultaneously to capture richer latent correlations results in redundant computations, thereby increasing the scale of the network and raising computational costs. To address these challenges, we propose a digital rock image super-resolution reconstruction network incorporating a Lightweight Two-Dimensional Multi-Scale Large Kernel Attention Building Block (LDMB). The LDMB consists of two Metaformer-Style Convolution Building Blocks (MSCB), along with two spatial attention units and channel attention units. The MSCB consists of a Multi-Scale Large Kernel Attention Unit (MLKA), and an innovative Gated Depthwise Convolutional Feedforward Neural Network (GCFN). Notably, we design a combined use of spatial and channel attention by innovatively proposing Enhanced Dense Convolutional Matrix Spatial Attention (EDSA) and Optimized Blueprint Contrast-Aware Channel Attention (OBCA), which integrates channel information from a global perspective, effectively enhancing the model’s ability to capture subtle textures and complex pore structures in digital rock images. Finally, to improve the contrast of the reconstructed rock images, we employ threshold segmentation techniques to process the images.