Scholarlink

<p>With the rapid advancement of deep learning, lightweight single image super-resolution has achieved remarkable progress. However, existing methods predominantly focus on deep feature extraction, while the effective utilization of shallow features remains insufficient. Efficient extraction of shallow features plays a crucial role in reconstruction accuracy. To address this, we propose the Efficient Shallow Feature Extraction Transformer (ESFET), which comprises the Gradient Vector Characterization Module (GVCM) and the Self-Referential Dual-Branch Transformer Module (SDTM). GVCM analyzes gradient variations in image textures to capture the transformation patterns of low-resolution (LR) images in the gradient vector space, generating divergence and curl images. This enables a refined representation of images in terms of numerical dispersion and directional rotation. SDTM adopts a dual-branch structure where the divergence and curl images serve as the keys in the Transformer and act as self-referential images, while the LR image functions as the queries and values. By computing correlations between the LR and reference images, SDTM leverages an attention mechanism to focus on texture fluctuations and integrates these fine-grained representations with conventional convolution-extracted features, thereby enhancing shallow feature extraction. Experimental results demonstrate that ESFET significantly improves reconstruction accuracy and surpasses state-of-the-art methods across multiple benchmark datasets, highlighting its superior performance.</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

Efficient shallow feature extraction for lightweight super-resolution via transformer

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Haoran Jia,
  • Tongtai Cao,
  • Xin Wang,
  • Pengjie Zhao,
  • Yue Liu

摘要

With the rapid advancement of deep learning, lightweight single image super-resolution has achieved remarkable progress. However, existing methods predominantly focus on deep feature extraction, while the effective utilization of shallow features remains insufficient. Efficient extraction of shallow features plays a crucial role in reconstruction accuracy. To address this, we propose the Efficient Shallow Feature Extraction Transformer (ESFET), which comprises the Gradient Vector Characterization Module (GVCM) and the Self-Referential Dual-Branch Transformer Module (SDTM). GVCM analyzes gradient variations in image textures to capture the transformation patterns of low-resolution (LR) images in the gradient vector space, generating divergence and curl images. This enables a refined representation of images in terms of numerical dispersion and directional rotation. SDTM adopts a dual-branch structure where the divergence and curl images serve as the keys in the Transformer and act as self-referential images, while the LR image functions as the queries and values. By computing correlations between the LR and reference images, SDTM leverages an attention mechanism to focus on texture fluctuations and integrates these fine-grained representations with conventional convolution-extracted features, thereby enhancing shallow feature extraction. Experimental results demonstrate that ESFET significantly improves reconstruction accuracy and surpasses state-of-the-art methods across multiple benchmark datasets, highlighting its superior performance.