Scholarlink

<p>Super-resolution reconstruction of remote sensing images is a critical research area in computer vision, with broad applications in environmental monitoring, disaster early warning, and military reconnaissance. Variable-resolution optical satellite images, generated by vertical orbit ring scanning sensors, exhibit high resolution at the sub-satellite point and gradually decreasing resolution toward the edge, leading to compromised image quality in edge regions and limiting overall data utilization. Traditional interpolation methods and mainstream super-resolution models, primarily designed for natural images, fail to effectively adapt to the spatial resolution variation of such remote sensing data, resulting in suboptimal reconstruction outcomes. To address this challenge, this paper proposes a Multi-Scale Feature Fusion Attention Network (MSFFAN) for edge region super-resolution of variable-resolution optical satellite images. First, a comprehensive degradation model simulating optical blurring, sensor sampling, and noise interference is established to generate a realistic training dataset. Second, the MSFFAN constructs multi-scale inputs via pixel unshuffling, extracts features using residual dense networks, and integrates shallow spatial and deep semantic information. Finally, a feature enhancement module combining group architecture and dual-attention mechanisms dynamically adjusts channel weights and focuses on high-frequency detail regions. Here we show that compared with the widely used Real-ESRGAN model, the MSFFAN achieves a PSNR improvement of 2.35 dB and an SSIM increase of 0.04, with significantly enhanced texture details in reconstructed images. This research provides an effective technical solution for improving the edge quality of variable-resolution remote sensing images, contributing to the broader application of such data in key fields and advancing the development of super-resolution technology for remote sensing. Code can be available at <a href="https://github.com/yloveyu/MSFFAN---master">https://github.com/yloveyu/MSFFAN---master</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

Multi-scale feature fusion attention network: enhancing edge super-resolution of variable-resolution optical satellite images

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Yu Qiao,
  • Peng Wang,
  • Wu Xue,
  • Xian Liu

摘要

Super-resolution reconstruction of remote sensing images is a critical research area in computer vision, with broad applications in environmental monitoring, disaster early warning, and military reconnaissance. Variable-resolution optical satellite images, generated by vertical orbit ring scanning sensors, exhibit high resolution at the sub-satellite point and gradually decreasing resolution toward the edge, leading to compromised image quality in edge regions and limiting overall data utilization. Traditional interpolation methods and mainstream super-resolution models, primarily designed for natural images, fail to effectively adapt to the spatial resolution variation of such remote sensing data, resulting in suboptimal reconstruction outcomes. To address this challenge, this paper proposes a Multi-Scale Feature Fusion Attention Network (MSFFAN) for edge region super-resolution of variable-resolution optical satellite images. First, a comprehensive degradation model simulating optical blurring, sensor sampling, and noise interference is established to generate a realistic training dataset. Second, the MSFFAN constructs multi-scale inputs via pixel unshuffling, extracts features using residual dense networks, and integrates shallow spatial and deep semantic information. Finally, a feature enhancement module combining group architecture and dual-attention mechanisms dynamically adjusts channel weights and focuses on high-frequency detail regions. Here we show that compared with the widely used Real-ESRGAN model, the MSFFAN achieves a PSNR improvement of 2.35 dB and an SSIM increase of 0.04, with significantly enhanced texture details in reconstructed images. This research provides an effective technical solution for improving the edge quality of variable-resolution remote sensing images, contributing to the broader application of such data in key fields and advancing the development of super-resolution technology for remote sensing. Code can be available at https://github.com/yloveyu/MSFFAN---master.