中国空间科学技术 ›› 2021, Vol. 41 ›› Issue (3): 16-23.doi: 10.16708/j.cnki.1000.758X.2021.0033

• 研究探讨 • 上一篇    下一篇

基于生成对抗网络的空间卫星低照度图像增强

陈榆琅1,高晶敏1,*,张科备2,张洋1   

  1. 1 北京信息科技大学自动化学院,北京100192
    2 北京控制工程研究所,北京100190
  • 收稿日期:2020-07-28 修回日期:2020-09-08 接受日期:2020-09-15 出版日期:2021-06-25 发布日期:2021-06-25
  • 通讯作者: Email: gaojm_biti@163.com E-mail:gaojm_biti@163.com
  • 作者简介:陈榆琅(1995-),男,研究生,研究方向为卫星图像增强及局部构件检测,2018020261@mail.bistu.edu.cn。 高晶敏(1966-),女,教授,研究方向为智能检测技术、多信息融合与处理,gaojm_biti@163.com。
  • 基金资助:
    国防科工局稳定支持项目(HTKJ2019KL502008);“十三五”民用航天技术预先研究项目(D020103、D030105)

Low-light image enhancement of space satellites based on GAN

CHEN Yulang1,GAO Jingmin1,*,ZHANG Kebei2,ZHANG Yang1   

  1. 1 School of Automation, Beijing Information Science& Technology University, Beijing 100192, China
    2 Beijing Institute of Control Engineering, Beijing 100190, China
  • Received:2020-07-28 Revised:2020-09-08 Accepted:2020-09-15 Published:2021-06-25 Online:2021-06-25
  • Contact: Email: gaojm_biti@163.com E-mail:gaojm_biti@163.com

摘要: 针对空间低照度成像条件下卫星光学图像信息受损严重的问题,提出了一种基于生成对抗网络的空间卫星低照度图像增强方法,提高了图像的平均亮度及对比度,恢复图像细节信息,为图像识别等图像处理技术提供更高质量的数据信息。首先,设计了一种密集连接的生成器,加强了各特征提取阶段中的信息传递以及多层特征的融合,减少了特征信息的损耗,更好地提取正常照度图像及低照度图像中相似的语义信息;并结合EnlightenGAN的思想,采用了全局局部辨别器结构,使图像增强效果更自然。然后,在少量样本的条件下,利用非配对样本对该方法进行训练,并通过对输入图像进行随机缩放及翻转等数据增强方法提高模型训练效果,进而提升低照度图像增强性能。最后,对所提出的空间卫星低照度图像增强方法进行了仿真验证。试验结果表明,在空间低照度条件下,该方法在NIQE指标上较LIME及EnlightenGAN分别降低了1.034和0.699,保留了更多的图像细节,具有更高的整体和局部亮度、更高的对比度以及更自然的增强效果。

关键词: 低照度图像增强, 生成对抗网络, 非配对训练, 密集连接, 相对辨别器

Abstract: Aiming at the problem of serious information damage of satellite optical images under the lowlight imaging condition, we proposed a satellite lowlight image enhancement method based on GAN. The method can improve the average brightness and contrast of images, restore image details, and provide higherquality information for image processing techniques such as image recognition. Firstly, we designed a densely connected generator to strengthen the information propagation and fusion between each feature extraction phase, reduce the loss of feature, and better extract similar semantic information in normallight and lowlight images. Combining the idea of EnlightenGAN, the globallocal discriminator structure was introduced to enhance images more naturally. Under the condition of a small number of samples, unpaired training was used to the proposed method, and data enhancement methods such as random scaling and flipping of the input images were applied to improve the training effect and model performance. Finally, the proposed method was validated by simulation. The experimental results show that, under the condition of low illumination, the proposed method reduced NIQE by 1.034 and 0.699 compared with the LIME and EnlightenGAN. The proposed method can preserve more image details, realize higher overall and local brightness, higher contrast, and more natural effects of enhancement.

Key words: low-light image enhancement, GAN, unpaired training, dense connection, relativistic discriminator

中图分类号: