Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (13): 190-199.DOI: 10.3778/j.issn.1002-8331.2304-0200

• Graphics and Image Processing • Previous Articles     Next Articles

Remote Sensing Image Super-Resolution Reconstruction Method for Ship Target Recognition

ZHANG Tianlin, PANG Zheng, CHEN Hongzhen, CHEN Shi, BIAN Chunjiang   

  1. 1.Key Laboratory of Electronics and Information Technology for Space Systems, National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China
    2.School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing 100049, China
  • Online:2024-07-01 Published:2024-07-01

面向舰船目标识别的遥感图像超分辨率重建

张天霖,逄征,陈红珍,陈实,卞春江   

  1. 1.中国科学院 国家空间科学中心 复杂航天系统综合电子与信息技术重点实验室,北京 100190
    2.中国科学院大学 计算机科学与技术学院,北京 100049

Abstract: The degradation of space-based remote sensing image resolution poses great challenges for the recognition of ship targets. Image super-resolution reconstruction technology can provide rich information for recognition tasks. However, if image super-resolution reconstruction and ship target recognition tasks are performed independently, the internal coherence between the two tasks will be ignored. Aiming at these problems, in order to explore the effective combination of image super-resolution reconstruction and target recognition tasks, a remote sensing image super-resolution reconstruction method for ship target recognition is proposed. Specifically, a full channel concatenation network is firstly designed, which replaces the residual connection with an adaptively weighted full channel concatenation, improves the fluidity and expression performance of each layer feature, and realizes efficient super-resolution reconstruction of remote sensing images. In order to further explore the potential of super-resolution reconstruction to improve the performance of ship target recognition, a joint network of super-resolution reconstruction and target recognition is proposed by introducing multi-task learning technology. The stable training of the joint end-to-end network is realized by multi-stage training optimization strategy, so as to guide effective mutual supervised learning between tasks. The experimental results on the public data set FGSCR-42 show that when the resolution of remote sensing images is degraded by 8 times and 16 times, the proposed super-resolution reconstruction network helps the accuracy of ship target recognition to increase by 33.27 and 17.48 percentage points respectively; the proposed joint network further improves the recognition accuracy by 1.75 and 1.91 percentage points.

Key words: space-based remote sensing images, image super-resolution reconstruction, ship target recognition

摘要: 天基遥感图像分辨率的退化为舰船目标的识别带来了巨大挑战。图像超分辨率重建技术可以为识别任务提供丰富的信息,然而将图像超分辨率重建与舰船目标识别任务分别独立进行,将会忽略两个任务间的内在相关性。针对这些问题,为了探索图像超分辨率重建与目标识别任务间的有效结合方式,提出了面向舰船目标识别的遥感图像超分辨率重建方法。设计了一种通道全连接网络,以自适应加权的通道全连接代替残差连接,提升各层特征的流动性与表达性能,实现遥感图像的高效超分辨率重建。为了进一步挖掘超分辨率重建对舰船目标识别性能提升的潜力,引入多任务学习技术,提出了一种超分辨率重建与目标识别联合网络,通过多阶段训练优化策略实现联合端到端网络的稳定训练,从而引导任务间进行有效的互相监督学习。在公开数据集FGSCR-42上的实验结果表明,当遥感图像在8倍和16倍的分辨率退化情况下,提出的超分辨率重建网络帮助舰船目标识别准确率分别提升了33.27和17.48个百分点,所提联合网络则将识别准确率进一步提升了1.75和1.91个百分点。

关键词: 天基遥感图像, 图像超分辨重建, 舰船目标识别