摘要
针对目前深度卷积神经网络(Convolutional Neural Network,CNN)在遥感图像建筑物提取上存在小目标漏分、被遮挡目标无法提取、细节缺失等问题,在生成对抗网络(Generative Adversarial Network,GAN)的基础上提出一种基于多尺度条件生成对抗网络(Multi-Scale Conditional Generative Adversarial Network,MSR-cGAN)的城市建筑物提取方法.该方法包括生成网络和对抗网络两个部分,在生成网络中加入循环残差卷积模块和注意力门限跳跃连接机制,增强模型的特征提取能力;在对抗网络中引入通道注意力的特征融合,使网络提取丰富的上下文信息,应对目标尺度变化,改善小目标分割效果.在实验过程中,对Inria Aerial Image Labeling建筑物提取数据集进行实验并与多种方法进行比较,结果表明,所提出的方法具有更高的目标分割准确率,对小目标与被遮挡目标取得了较好的分割效果.在训练数据有限、背景复杂多样、尺度变化较大的建筑物提取中分割准确率分别达到96.18%,表明提出的方法可应用于复杂的高分辨率遥感图像建筑物提取.
At present,there are some problems existing in deep convolutional neural network(CNN)in building extraction of remote sensing image,such as missing small targets,unable to extract occluded targets,and missing details. On the basis of Generative Adversarial Network(GAN),a method for urban building extraction based on Multi-Scale Conditional Generative Adversarial Network(MSR-cGAN)is proposed. The method includes generative network and adversarial network. In the process of the experiment,the Inria Aerial Image Labeling building extraction data set is tested and compared with a variety of methods. The results show that the proposed method has higher target segmentation accuracy,and has better segmentation results for small targets and occluded targets. The segmentation accuracy reaches 96.18% in building extraction with limited training data,complex background and large scale change,respectively. It shows that the proposed method can be applied to complex high-resolution remote sensing image building extraction.
作者
郭杨亮
马瑞娟
韩子清
GUO Yangliang;MA Ruijuan;HAN Ziqing(Henan Institute of Geophysical and Spatial Information,Zhengzhou 450009,China;Fifth Geological Exploration Institute of Henan Geological and Mineral Exploration and Development Bureau,Zhengzhou 450052,China;Henan Geological Survey Institute,Zhengzhou 450007,China)
出处
《河南科学》
2022年第9期1377-1383,共7页
Henan Science
关键词
卷积神经网络
生成对抗网络
高分辨率遥感图像
建筑物提取
convolutional neural network
generative adversarial network
high-resolution remote sensing image
building extraction
