Scholarlink

<p>In the field of image super-resolution, the challenge of achieving high-fidelity image recovery poses a significant obstacle for shallow convolutional neural networks (CNNs). While deepening or widening CNNs can improve both perceptual and numerical results to some extent, the corresponding increase in computational complexity remains impractical for real-world applications. To tackle this issue, we propose a lightweight multi-semantic pixel attention network aimed at overcoming the performance limitations of shallow networks. In particular, our proposed multi-semantic pixel attention (MSPA) mechanism leverages rich semantic information to render more realistic textures at a low computational cost. Building upon the concepts of dense connections and information distillation, we devise an efficient information distillation dense block (IDDB). The IDDB is created by alternately stacking adjustment units (AUs) and correction units (CUs). AUs are responsible for receiving and nonlinearly fusing features from preceding layers. The fused features are then split into two parts: one flows to subsequent AUs to enrich information, while the other proceeds to CUs for further refinement. The incorporation of MSPA further enhances the performance of the IDDB. Comprehensive experiments demonstrate that our approach achieves superior quantitative metrics and visual quality while maintaining lower model complexity compared to other state-of-the-art models. Our code will be made available at: <a href="https://github.com/cipher-wzy/MSPAN">https://github.com/cipher-wzy/MSPAN</a>.</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

MSPAN: lightweight image super-resolution with multi-semantic guidance

  • 期刊论文

  • 发表日期：

    2025年1月6日

查看全文
  • Zheyuan Wang,
  • Ziyao Meng,
  • Yiming Qin

摘要

In the field of image super-resolution, the challenge of achieving high-fidelity image recovery poses a significant obstacle for shallow convolutional neural networks (CNNs). While deepening or widening CNNs can improve both perceptual and numerical results to some extent, the corresponding increase in computational complexity remains impractical for real-world applications. To tackle this issue, we propose a lightweight multi-semantic pixel attention network aimed at overcoming the performance limitations of shallow networks. In particular, our proposed multi-semantic pixel attention (MSPA) mechanism leverages rich semantic information to render more realistic textures at a low computational cost. Building upon the concepts of dense connections and information distillation, we devise an efficient information distillation dense block (IDDB). The IDDB is created by alternately stacking adjustment units (AUs) and correction units (CUs). AUs are responsible for receiving and nonlinearly fusing features from preceding layers. The fused features are then split into two parts: one flows to subsequent AUs to enrich information, while the other proceeds to CUs for further refinement. The incorporation of MSPA further enhances the performance of the IDDB. Comprehensive experiments demonstrate that our approach achieves superior quantitative metrics and visual quality while maintaining lower model complexity compared to other state-of-the-art models. Our code will be made available at: https://github.com/cipher-wzy/MSPAN.