测绘学报 ›› 2023, Vol. 52 ›› Issue (4): 648-659.doi: 10.11947/j.AGCS.2023.20210571

• 摄影测量学与遥感 • 上一篇    下一篇

顾及全局特征和纹理特征的遥感影像超分辨率重建方法

胡安娜1, 刘睿2, 吴亮3, 张进4, 徐永洋3, 陈思琼2   

  1. 1. 中国地质大学(武汉)国家地理与信息系统工程技术研究中心, 武汉 430078;
    2. 中国地质大学(武汉)地理与信息工程学院, 武汉 430078;
    3. 中国地质大学(武汉)计算机学院, 武汉 430078;
    4. 武汉科技大学计算机科学与技术学院, 武汉 430081
  • 收稿日期:2021-10-20 修回日期:2022-05-10 发布日期:2023-05-05
  • 通讯作者: 徐永洋 E-mail:yongyangxu@cug.edu.cn
  • 作者简介:胡安娜(1994-),女,博士生,研究方向为遥感影像处理。E-mail:huanna@cug.edu.cn
  • 基金资助:
    国家自然科学基金(41871311;42001340)

Super-resolution reconstruction method for remote sensing images considering global features and texture features

HU Anna1, LIU Rui2, WU Liang3, ZHANG Jin4, XU Yongyang3, CHEN Siqiong2   

  1. 1. National Engineering Research Center of Geographic Information System, China University of Geosciences, Wuhan 430078, China;
    2. School of Geography and Information Engineering, China University of Geosciences, Wuhan 430078, China;
    3. School of Computer Science, China University of Geosciences, Wuhan 430078, China;
    4. School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan 430081, China
  • Received:2021-10-20 Revised:2022-05-10 Published:2023-05-05
  • Supported by:
    The National Natural Science Foundation of China (Nos. 41871311;42001340)

摘要: 由于遥感设备的性能限制,使得采集的遥感影像质量受到影响,低分辨率的遥感影像限制了遥感解译应用的精度。当前针对遥感影像的超分辨率重建研究仍然存在重建后的遥感影像地物全局信息和纹理细节不足的问题。因此,本文提出顾及全局特征和纹理特征的遥感影像超分辨率重建方法,该方法利用生成对抗网络的特征学习能力,并对模型全局和纹理进行增强。一方面,地物全局特征增强部分用于解决当前研究中超分辨率重建模型对低分辨率遥感影像中全局遥感地物信息没有重视和利用的问题。在生成网络中引入自注意力模块,以获取全局地物注意力图的方式将遥感影像中相距较远的地物信息作为重建过程的参考。另一方面,遥感影像纹理增强部分用于解决超分辨率重建模型中超分辨率影像纹理信息不足的问题。本文方法引入纹理损失以优化生成网络参数并增强超分辨率重建后影像中的纹理信息。另外,为避免重建结果中的“伪影”现象,研究采用权值归一化代替批量归一化方法。试验结果表明,本文方法在遥感影像超分辨率重建过程中能增强遥感地物特征,同时可以实现地物的纹理细节精细化恢复,而且超分辨率重建结果的图像质量评价指标SSIM、FSIM和PSNR值分别达到了0.756、0.595和26.005。

关键词: 高分辨率遥感影像, 超分辨率重建, 生成对抗网络, 纹理增强, 全局特征

Abstract: Due to the performance limitation of remote sensing equipment, the quality of the remote sensing image is affected, and the low-resolution remote sensing image limits the accuracy of remote sensing interpretation applications. Insufficient global information and texture details of reconstructed remote sensing images are still in super-resolution reconstruction research. Therefore, this study proposes a super-resolution reconstruction method for remote sensing images considering global features and texture features. The method utilizes the feature learning ability of the generative adversarial network to optimize the model in two aspects: global information enhancement and texture information enhancement. On the one hand, the global feature enhancement part is used to solve the problem that the current super-resolution reconstruction model does not pay attention to the global remote sensing information of low-resolution remote sensing images. The self-attention module is introduced to the generation network, which is used to obtain the global object attention map, and the remote object information in remote sensing image is used as a reference in the reconstruction process. On the other hand, the texture enhancement part is used to solve the problem of insufficient texture information of the reconstructed remote sensing image. Texture loss is introduced to the optimized generated network that the texture information of ground objects can be improved. In addition, weight normalization is adopted to replace batch normalization to avoid false shadows in the reconstruction result. The experimental results show that the proposed super-resolution algorithm can not only enhance the features of the ground object, but also recovery the texture details for ground objects, and the SSIM, FSIM, and PSNR values of the reconstructed super-resolution image quality evaluation index are 0.756, 0.595 and 26.005, respectively.

Key words: super-resolution reconstruction, high-resolution remote sensing image, generative adversarial networks, texture enhancement, global feature

中图分类号: