多模态融合的高分遥感图像语义分割方法
Multi-modal Fusion Based Method for High Resolution Remote Sensing Image Segmentation
投稿时间:2019-12-26  修订日期:2019-12-26
DOI:
中文关键词: 遥感图像分割  卷积神经网络  DSM  U-Net
英文关键词: remote sensing  image segmentation  CNN  DSM
基金项目:中南民族大学中央高校基本科研业务费专项资金项目(CZY18002)
作者单位E-mail
陈少波 中南民族大学电子与信息工程学院 chenshaobo1980@qq.com 
李万琦 中南民族大学电子与信息工程学院 liwanqi317@126.com 
李克俭 中南民族大学电子与信息工程学院  
摘要点击次数: 73
全文下载次数: 0
中文摘要:
      遥感图像的语义分割是遥感图像解译的基础,卫星在俯拍时造成的空间距离信息丢失,使得高分辨率遥感图像呈现出地物信息复杂、目标尺度多样化,类间差距小等特点,这些特点使得遥感图像的语义分割极具挑战。目前,以深度神经网为基础的遥感图像语义分割取得了一定的进展,这类方法主要利用深度网络训练RGB图像得到特征图用于像素点的识别分类。然而俯拍图像固有的对高度信息不敏感或缺乏,导致传统模型提取的特征缺乏景深信息。为解决上述问题,本文提出一种将数值地表模型(Digital Surface Model, DSM)与RGB信息相结合的多模态融合模型SEU-Net(Squeeze and Extraction U-Net),该模型使用SE模块将DSM图像的空间信息编码为权重加至对应的RGB图像的特征图通道上,以获得更具判别性的特征,并以端到端的方式输出分割图片。本文提出的模型在ISPRS遥感图像分割数据集上进行了实验验证,取得了88.8%的分割精度;对高度信息敏感的目标,分割效果提升明显。
英文摘要:
      Semantic segmentation of RS (Remote Sensing) images is basis of understanding RS images. One major difficulty is that the satellite photos taken from the above can cause the loss of spatial distance information, thus, the targets vary in different scales and have small differences in between. Nowadays, some progress has been made in this task based on deep neural networks by training the RGB images through the network to obtain the feature maps and using them to get segmentation result. However, the overhead images are inherently insensitive or lacking in height information, resulting in features extracted from traditional models lacking this sort of information. In order to solve the above problem, this paper proposes a multi-modal fusion model SEU-Net (Squeeze and Extraction U-Net) , which combining digital surface model (DSM) and RGB information. The height information contains in DSM images is encoded via SE module as weights added to the channel corresponding to the feature map of RGB images, the classification of each pixel is predicted in an end-to-end manner. The model proposed in this paper has been experimentally verified on the ISPRS remote sensing image segmentation dataset, and has achieved overall accuracy of 88.8%, also it has better segmentation results for targets that are sensitive to height.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭