Scholarlink

<p>With the explosive growth in the number of spacecrafts in space, space target measurement tasks have become an indispensable strategic component of space missions. High-resolution (HR) images are necessary for these space measurement tasks such as rendezvous and docking, interstellar observations and manned spaceflight. However, due to the high speed and manoeuvrability of the targets, limitation in on-board equipment, and the complex space environment, the acquired space target observation images suffer from blurring, artifacts, and loss of detail, which limits the accuracy of space measurement missions, seriously. In recent years, with the advancement of computer technology and deep learning, an end-to-end image restoring technique named super-resolution reconstruction (SR) has developed. Therefore, in this paper, a lightweight SR method called enhanced adaptive kernel selection network (EAKSN) is proposed to improve the quality of images in space measurement tasks with little calculation and time consumption. Several state-of-the-art attention modules and blocks, such as adaptive feature extraction large kernel attention (AFELA), enhanced multi-scale selective kernel module (EMSKA(M)), and enhanced adaptive feature distillation block (EAFDB) are designed in EAKSN. Aiming at adapting satellite devices, this method is lightweight enough by introducing blueprint separable convolution (BS-Conv) and CNN based attention modules, which allows this method to have a very small number of parameters and flops, fast running-time, and powerful reconstruction ability. Finally, a space target observation dataset (STO-NWPU) is designed to test the performance of our method in addition to public datasets like Set5, Set14, Urban100 and BSD100. Experimental results demonstrate that EAKSN achieves up to 0.49&#xa0;dB PSNR improvement over representative lightweight super-resolution methods on Urban100 at × 4 under bicubic degradation, while maintaining fewer than 0.7&#xa0;M parameters and low computational complexity making it suitable for deploying on satellites.</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

A lightweight CNN-based method to improve the quality of images in space measurement tasks

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Bingzan Liu,
  • Min Zhu,
  • Shuli Liang,
  • Kouan Hao,
  • Yizhen Yang,
  • Shengjie Tian

摘要

With the explosive growth in the number of spacecrafts in space, space target measurement tasks have become an indispensable strategic component of space missions. High-resolution (HR) images are necessary for these space measurement tasks such as rendezvous and docking, interstellar observations and manned spaceflight. However, due to the high speed and manoeuvrability of the targets, limitation in on-board equipment, and the complex space environment, the acquired space target observation images suffer from blurring, artifacts, and loss of detail, which limits the accuracy of space measurement missions, seriously. In recent years, with the advancement of computer technology and deep learning, an end-to-end image restoring technique named super-resolution reconstruction (SR) has developed. Therefore, in this paper, a lightweight SR method called enhanced adaptive kernel selection network (EAKSN) is proposed to improve the quality of images in space measurement tasks with little calculation and time consumption. Several state-of-the-art attention modules and blocks, such as adaptive feature extraction large kernel attention (AFELA), enhanced multi-scale selective kernel module (EMSKA(M)), and enhanced adaptive feature distillation block (EAFDB) are designed in EAKSN. Aiming at adapting satellite devices, this method is lightweight enough by introducing blueprint separable convolution (BS-Conv) and CNN based attention modules, which allows this method to have a very small number of parameters and flops, fast running-time, and powerful reconstruction ability. Finally, a space target observation dataset (STO-NWPU) is designed to test the performance of our method in addition to public datasets like Set5, Set14, Urban100 and BSD100. Experimental results demonstrate that EAKSN achieves up to 0.49 dB PSNR improvement over representative lightweight super-resolution methods on Urban100 at × 4 under bicubic degradation, while maintaining fewer than 0.7 M parameters and low computational complexity making it suitable for deploying on satellites.