计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (21): 210-217.DOI: 10.3778/j.issn.1002-8331.2002-0202

• 图形图像处理 • 上一篇    下一篇

多尺度遥感语义分割网络

胥智杰,杨小兵,何灵敏,潘承瑞   

  1. 1.中国计量大学 信息工程学院,杭州 310000
    2.中国计量大学 浙江省电磁波信息技术与计量检测重点实验室,杭州 310000
  • 出版日期:2020-11-01 发布日期:2020-11-03

Multiscale Remote Sensing Semantic Segmentation Network

XU Zhijie, YANG Xiaobing, HE Lingmin, PAN Chengrui   

  1. 1.College of Information Engineering, China Jiliang University, Hangzhou 310000, China
    2.Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province, China Jiliang University, Hangzhou 310000, China
  • Online:2020-11-01 Published:2020-11-03

摘要:

高分辨率遥感图像语义分割在国土规划、地理监测、智慧城市等领域有着广泛的应用价值,但是现阶段研究中存在相似地物和精细地物分割不准确问题。为解决这一问题,提出了一种新型的多尺度语义分割网络MSSNet。它由编码层、解码层和输出层组成。为解决相似地物的分割问题,编码层使用深层网络ResNet101充分提取地物特征,并在解码层的解码器中加入残差块,提高基于像素点的分类能力。为解决精细结构地物的分割问题,解码层中的解码器加入了空洞空间金字塔池化结构提取多尺度地物特征,以便精确分割不同尺度的地物。为了强化语义分割能力,输出层合并了多个解码器的输出,为最终的预测提供了更多的信息。在两个公开数据集Vaihingen和Potsdam上进行了实验,分别取得了87%和87.3%的全局精确度,超过了大多数已发表的方法。实验结果表明,提出的MSSNet能够精确地分割相似地物和精细地物,并且具有训练过程简单和易于使用的优点,非常适合进行高分辨率遥感图像语义分割。

关键词: 高分辨率遥感图像, 语义分割, 深度学习, 多尺度语义分割网络(MSSNet)

Abstract:

The semantic segmentation of very-high-resolution remote sensing images have wide application value in land planning, geographic monitoring and smart cities. But at present study, there are the problems of inaccurate segmentation of similar objects and fine-structured objects. In order to solve these problems, a new Multiscale Semantic Segmentation Network(MSSNet) is proposed. It consists of encoding layer, decoding layer and output layer. For the problem of similar objects, ResNet101 is used in the encoding layer to fully extract object features and residual block is added in the decoder of the decoding layer to improve the classification ability based on pixel points. For problem of fine-structured objects segmentation, the atrous spatial pyramid pooling is added in the decoder to extract multi-scale object features. So that objects of different scales can be accurately segmented. To enhance the ability of semantic segmentation, the output layer combines the output of multiple decoders. This provides more information for the final forecast. Experiments on two open datasets, Vaihingen and Potsdam, have achieved overall accuracy of 87% and 87.3% respectively, which surpasses most published methods. The results show that MSSNet can accurately segment similar objects and fine-structured objects. And it has the advantages of simple training process and easy to use. It is very suitable for semantic segmentation of very-high-resolution remote sensing images.

Key words: very-high-resolution remote sensing image, semantic segmentation, deep learning, Multiscale Semantic Segment Network(MSSNet)