TY - GEN
T1 - A lightweight super-resolution reconstruction method for low-light space target observation images
AU - Liu, Bingzan
AU - Ning, Xin
AU - Ma, Shichao
AU - Chen, Hongyu
AU - Yang, Yizhen
N1 - Publisher Copyright:
Copyright ©2024 by the International Astronautical Federation (IAF). All rights reserved.
PY - 2024
Y1 - 2024
N2 - With the development in space science and technology, space target observation tasks have become an indispensable component of space missions and acquiring high-resolution (HR) images is the key for accomplishing target observation tasks. However, obtaining HR space target observation images directly is challenging due to the extreme light environment in space, limitations of on-board equipment, and high-speed movement of space targets. In recent years, with the development of super-resolution (SR) reconstruction methods, particularly the emergence of SR reconstruction technology based on deep learning, it has become possible to recover more detailed textures and high-frequency information from low-resolution (LR) images. However, researches on SR methods for low-light and unevenly illuminated spatial environments are still relatively scarce. Noteworthily, most of these methods have large number of parameters, flops, and running-time, which are not appropriate for on-board computers and space target observation tasks. Furthermore, due to the varying spatial illumination, it is difficult to achieve sample pairs of high- and low-quality images. Therefore, a lightweight SR reconstruction method for zero-shot space target observation images named GNN-based chain multi-scale adaptive fusion network (GCMAFN) is proposed. This network is composed by several state-of-the-art attention modules and blocks such as GNN-based chain multi-scale adaptive fusion block, multi-scale adaptive fusion block (MAFB). Additionally, aiming at evaluating the performance of our method, a space target observation images dataset (STO-NWPU) is built by us. Experimental results on benchmark datasets (Set5, Urban100, BSD100) and self-built STO-NWPU demonstrate that our GCMAFN outperforms in performance, parameters and running-time in low-light images and can be deployed on spacecrafts for improving the quality of target observation tasks.
AB - With the development in space science and technology, space target observation tasks have become an indispensable component of space missions and acquiring high-resolution (HR) images is the key for accomplishing target observation tasks. However, obtaining HR space target observation images directly is challenging due to the extreme light environment in space, limitations of on-board equipment, and high-speed movement of space targets. In recent years, with the development of super-resolution (SR) reconstruction methods, particularly the emergence of SR reconstruction technology based on deep learning, it has become possible to recover more detailed textures and high-frequency information from low-resolution (LR) images. However, researches on SR methods for low-light and unevenly illuminated spatial environments are still relatively scarce. Noteworthily, most of these methods have large number of parameters, flops, and running-time, which are not appropriate for on-board computers and space target observation tasks. Furthermore, due to the varying spatial illumination, it is difficult to achieve sample pairs of high- and low-quality images. Therefore, a lightweight SR reconstruction method for zero-shot space target observation images named GNN-based chain multi-scale adaptive fusion network (GCMAFN) is proposed. This network is composed by several state-of-the-art attention modules and blocks such as GNN-based chain multi-scale adaptive fusion block, multi-scale adaptive fusion block (MAFB). Additionally, aiming at evaluating the performance of our method, a space target observation images dataset (STO-NWPU) is built by us. Experimental results on benchmark datasets (Set5, Urban100, BSD100) and self-built STO-NWPU demonstrate that our GCMAFN outperforms in performance, parameters and running-time in low-light images and can be deployed on spacecrafts for improving the quality of target observation tasks.
KW - GNN-based chain multi-scale adaptive fusion network
KW - lightweight super-resolution reconstruction
KW - space target observation image
UR - https://www.scopus.com/pages/publications/105046150294
U2 - 10.52202/078362-0113
DO - 10.52202/078362-0113
M3 - 会议稿件
AN - SCOPUS:105046150294
T3 - Proceedings of the International Astronautical Congress, IAC
SP - 877
EP - 884
BT - IAF Human Spaceflight Symposium - Held at the 75th International Astronautical Congress, IAC 2024
PB - International Astronautical Federation, IAF
T2 - 2024 IAF Earth Observation Symposium at the 75th International Astronautical Congress, IAC 2024
Y2 - 14 October 2024 through 18 October 2024
ER -