Scholarlink

<p>With the advancement of astronomy, the demand for high-resolution astronomical images has become increasingly urgent. However, the imaging of ground-based telescopes is often affected by factors such as hardware limitations, atmospheric turbulence, and noise, thereby limiting the execution of high-precision scientific tasks. Moreover, due to the scarcity of paired high- and low-resolution observations, directly training super-resolution models on real data remains a significant challenge. Therefore, we construct a synthetic dataset, GalaxySR-DS, by simulating the degradation process using DownSampleGAN (DSGAN). In addition, we propose a lightweight super-resolution (SR) model, Frequency and Spatial-Channel Cross Transformer (FSCFormer). It combines the Frequency Transformation Module (FTB) and the Spatial-Channel Cross Transformation Module (SCCTB), and has complementary advantages in modeling high-frequency details and maintaining the global structure, thereby improving the reconstruction quality. We select ten representative lightweight SR models as baselines for comparison. Experimental results show that the FSCFormer model achieves an average PSNR improvement of 0.02-0.85 dB while maintaining lower computational cost. This study offers an effective approach for the high-precision analysis of astronomical observation data. In particular, under constrained observational resources, it maximizes the scientific value of existing data and promotes a paradigm shift in observational astronomy research. The code will be available at <a href="https://github.com/jiaweimmiao/FSCFormer">https://github.com/jiaweimmiao/FSCFormer</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

Frequency and spatial-channel cross transformer for lightweight astronomical image super-resolution

  • 期刊论文

  • 发表日期：

    2025年12月26日

查看全文
  • Jiawei Miao,
  • Liangping Tu,
  • Hao Liu,
  • Jian Zhao,
  • Zhengpeng Li,
  • Kunyang Wu

摘要

With the advancement of astronomy, the demand for high-resolution astronomical images has become increasingly urgent. However, the imaging of ground-based telescopes is often affected by factors such as hardware limitations, atmospheric turbulence, and noise, thereby limiting the execution of high-precision scientific tasks. Moreover, due to the scarcity of paired high- and low-resolution observations, directly training super-resolution models on real data remains a significant challenge. Therefore, we construct a synthetic dataset, GalaxySR-DS, by simulating the degradation process using DownSampleGAN (DSGAN). In addition, we propose a lightweight super-resolution (SR) model, Frequency and Spatial-Channel Cross Transformer (FSCFormer). It combines the Frequency Transformation Module (FTB) and the Spatial-Channel Cross Transformation Module (SCCTB), and has complementary advantages in modeling high-frequency details and maintaining the global structure, thereby improving the reconstruction quality. We select ten representative lightweight SR models as baselines for comparison. Experimental results show that the FSCFormer model achieves an average PSNR improvement of 0.02-0.85 dB while maintaining lower computational cost. This study offers an effective approach for the high-precision analysis of astronomical observation data. In particular, under constrained observational resources, it maximizes the scientific value of existing data and promotes a paradigm shift in observational astronomy research. The code will be available at https://github.com/jiaweimmiao/FSCFormer.