Scholarlink

We propose a super-resolution method for satellite images that extends the SwinIR Transformer architecture. Our network introduces residual triple attention groups (RTAGs) and triple attention blocks (TABs), which combine three attention mechanisms to enhance the restoration of local textures and global structures. We perform knowledge distillation between models of similar scale to further improve performance. After being trained on 7,500 Google Earth images, our model outperformed RCAN, HAT, and SwinIR in terms of PSNR and SSIM.

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

TAT-SatSR: Triple Attention Transformer for Super Resolution of Satellite Image

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shun Matsumoto,
  • Tohru Kamiya

摘要

We propose a super-resolution method for satellite images that extends the SwinIR Transformer architecture. Our network introduces residual triple attention groups (RTAGs) and triple attention blocks (TABs), which combine three attention mechanisms to enhance the restoration of local textures and global structures. We perform knowledge distillation between models of similar scale to further improve performance. After being trained on 7,500 Google Earth images, our model outperformed RCAN, HAT, and SwinIR in terms of PSNR and SSIM.