基于自注意力机制和CycleGAN的高分三号ScanSAR图像的扇贝效应抑制
Descalloping of GF-3 ScanSAR image based on self-attention mechanism and CycleGAN
孙增国 1彭学俊 2刘慧霞 3陈卫荣 4王鑫鹏2
作者信息
- 1. 陕西师范大学计算机科学学院,陕西西安 710119;地理信息工程国家重点实验室,陕西西安 710054
- 2. 陕西师范大学计算机科学学院,陕西西安 710119
- 3. 南通大学电气工程学院,江苏南通 226019
- 4. 中国资源卫星应用中心,北京 100094
- 折叠
摘要
高分三号卫星是我国首颗分辨率达到1 m的C波段多极化合成孔径雷达(synthetic aperture radar,SAR)卫星,其中扫描式合成孔径雷达(scan synthetic aperture radar,ScanSAR)模式是高分三号卫星重要的工作模式之一,由于该模式的工作机制导致生成的图像可能发生扇贝效应,一般呈现为明暗相间的条纹.本文针对高分三号卫星ScanSAR模式下存在的扇贝效应,提出自注意力机制与循环一致对抗网络(cycle-consistent adversarial networks,CycleGAN)结合的模型对 Scan-SAR图像进行处理,从而抑制扇贝效应产生的条纹现象.本文所示方法与传统扇贝效应抑制方法和深度学习相关算法进行比较,并通过亮度均值、平均梯度等指标进行分析.实验结果表明,本文方法可以对高分三号ScanSAR图像存在的扇贝效应进行较好的处理,有效抑制图像的条纹现象,使得图像质量得到提升,具有较大的实用意义.
Abstract
GF-3 satellite is the first C-band multi-polarimetric synthetic aperture radar satellite with a space resolution up to 1 m in China,in which scan synthetic aperture radar(ScanSAR)is one of the im-portant mode of GF-3.The working mechanism of this mode results in the phenomena of serious nonuni-formity,generally showing bright and dark stripes,also known as scalloping.In view of scalloping in ScanSAR mode of GF-3,this paper proposes a model combining self-attention mechanism and cycle-con-sistent adversarial networks(CycleGAN),so as to perform descalloping.The proposed descalloping method is compared with traditional descalloping methods and deep learning related algorithms,and ana-lyzed by indicators such as brightness average and average gradient.The experimental results demon-strate that the proposed method in this paper can better complete descalloping in the GF-3 ScanSAR im-age,effectively suppress the stripes phenomenon of the image,and improve the image quality,which is of great practical significance.
关键词
高分三号/扫描式合成孔径雷达(ScanSAR)图像/扇贝效应/循环一致对抗网络(Cy-cleGAN)/自注意力机制Key words
GF-3/scan synthetic aperture radar(ScanSAR)image/scalloping/cycle-consistent adver-sarial networks(CycleGAN)/self-attention mechanism引用本文复制引用
基金项目
国家自然科学基金(61102163)
地理信息工程国家重点实验室基金(SKLGIE2019-M-3-5)
出版年
2023
NETL
NSTL
维普
万方数据