Three-Class Change Detection in Synthetic Aperture Radar Images Based on Deep Belief Network
Three-Class Change Detection in Synthetic Aperture Radar Images Based on Deep Belief Network
复制标题
DOI：
10.1007/978-3-662-49014-3_62
复制
发表时间：
2015-09
期刊：
影响因子：
--
通讯作者：
Q. Zhao;Maoguo Gong;Hao Li;Tao Zhan;Qian Wang
中科院分区：
文献类型：
--
作者：
Q. Zhao;Maoguo Gong;Hao Li;Tao Zhan;Qian Wang
提出了一种基于深度学习的合成孔径雷达图像（SAR）三类变化检测方法。在大多数文献中，图像变化检测是将比率图像分为变化类和不变类。然而，多时相SAR图像的后向散射值要么增加要么减少，因此进一步将变化区域分为正变化和负变化类是有意义的。我们通过深度学习实现了这种新颖的三类变化检测方法。对于给定的多时相图像，利用改进的对数比算子生成一幅反映相应像素间差异程度的差分图像。然后，我们建立了一个深度信念网络来分析差异图像，识别正变化像素，负变化像素和不变像素。
In this paper, we propose a novel three-class change detection approach for synthetic aperture radar images (SAR) based on deep learning. In most literatures, change detection in images is a method that classifies the ratio images into two parts: the changed and unchanged classes. However, multitemporal SAR images have either increase or decrease in the backscattering values, so it is significative to further classify the changed areas into the positive and negative changed classes. We accomplish this novel three-class change detection method through Deep Learning. Given the multitemporal images, a difference image which shows difference degrees between corresponding pixels is generated by modified log-ratio operator. Then, we establish a deep belief network to analyze the difference image and recognize the positive changed pixels, negative changed pixels and unchanged pixels.