Scholarlink

<p>With the rapid advancement of neural networks, image super-resolution models based on convolutional neural networks have significantly progressed, consistently enhancing image reconstruction quality and detail preservation capabilities. However, deploying these sophisticated models on resource-constrained edge devices presents substantial challenges with respect to computational resource allocation and data storage requirements. To address the limitations of large models with high parameter counts and computational demands that render them unsuitable for edge devices, we propose a Lightweight Shuffle Feature Fusion Network (LSFFN). Our approach introduces two key innovations. First, we design a lightweight multi-scale shallow feature extraction module that employs depthwise separable convolutions and channel-shuffle operations, effectively reducing the computational overhead associated with multi-scale feature extraction. Second, we implement a shuffle feature fusion structure that preserves image details while facilitating efficient feature fusion in a resource-efficient manner, thus mitigating the performance degradation commonly observed in lightweight SR models. Compared with state-of-the-art lightweight SR methods, our proposed method demonstrates better performance. For instance, LSFFN outperforms IFIN-S by over 0.09 dB in PSNR and 0.0033 in SSIM when evaluated on the Urban100 (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation>4) dataset. Comprehensive evaluations demonstrate that LSFFN achieves enhanced reconstruction fidelity while significantly reducing parameter count and computational requirements, effectively striking an optimal balance between processing efficiency and image quality performance in resource-constrained environments.</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

Single-image super-resolution via lightweight shuffle feature fusion network

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Aiying Guo,
  • Zijun Deng,
  • Jingjing Liu

摘要

With the rapid advancement of neural networks, image super-resolution models based on convolutional neural networks have significantly progressed, consistently enhancing image reconstruction quality and detail preservation capabilities. However, deploying these sophisticated models on resource-constrained edge devices presents substantial challenges with respect to computational resource allocation and data storage requirements. To address the limitations of large models with high parameter counts and computational demands that render them unsuitable for edge devices, we propose a Lightweight Shuffle Feature Fusion Network (LSFFN). Our approach introduces two key innovations. First, we design a lightweight multi-scale shallow feature extraction module that employs depthwise separable convolutions and channel-shuffle operations, effectively reducing the computational overhead associated with multi-scale feature extraction. Second, we implement a shuffle feature fusion structure that preserves image details while facilitating efficient feature fusion in a resource-efficient manner, thus mitigating the performance degradation commonly observed in lightweight SR models. Compared with state-of-the-art lightweight SR methods, our proposed method demonstrates better performance. For instance, LSFFN outperforms IFIN-S by over 0.09 dB in PSNR and 0.0033 in SSIM when evaluated on the Urban100 ( \(\times \) 4) dataset. Comprehensive evaluations demonstrate that LSFFN achieves enhanced reconstruction fidelity while significantly reducing parameter count and computational requirements, effectively striking an optimal balance between processing efficiency and image quality performance in resource-constrained environments.