作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2019, Vol. 45 ›› Issue (1): 239-245. doi: 10.19678/j.issn.1000-3428.0049109

• 图形图像处理 • 上一篇    下一篇

一种基于改进FCN的多光谱图像建筑物识别方法

张永梅1,付昊天1,孙海燕1,张睿2,陈立潮2,潘理虎2   

  1. 1.北方工业大学 计算机学院,北京 100144; 2.太原科技大学 计算机科学与技术学院,太原 030024
  • 收稿日期:2017-10-30 出版日期:2019-01-15 发布日期:2019-01-15
  • 作者简介:张永梅(1967—),女,教授、博士,主研方向为图像处理、人工智能;付昊天、孙海燕,硕士研究生;张睿,讲师、博士;陈立潮,教授、博士;潘理虎,副教授、博士。
  • 基金资助:

    国家自然科学基金(61371143);北方工业大学优势学科项目(XN044);太原科技大学博士科研启动基金(20162036)

A Building Recognition Method for Multispectral Image Based on Improved FCN

ZHANG Yongmei1,FU Haotian1,SUN Haiyan1,ZHANG Rui2,CHEN Lichao2,PAN Lihu2   

  1. 1.School of Computer Science,North China University of Technology,Beijing 100144,China; 2.School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China
  • Received:2017-10-30 Online:2019-01-15 Published:2019-01-15

摘要:

多光谱图像的建筑物目标在不同尺度下具有不同特征,利用传统全卷积神经网络(FCN)进行识别时精度较低。为此,提出一种基于改进FCN的多光谱图像建筑物识别方法。通过旋转图像进行训练集扩充,从网络的第1层~第12层提取图像在4个旋转角度和不同尺度下的低层特征,将其归一化为同样尺寸的图像后提取更高层特征,以实现对多光谱图像建筑物的精确识别。实验结果表明,相比传统FCN方法,该方法能够提高识别的精确率与召回率。

关键词: 多光谱图像, 建筑物识别, 全卷积神经网络, 多尺度信息, 训练集扩充

Abstract:

Building targets in multispectral image have different characteristics at different scales,and the recognition accuracy of traditional Full Convolution Neural network(FCN) is low.Therefore,a building recognition method for multispectral image based on improved FCN is proposed.After expanding the training set by rotating images,the low-level features of images at four rotating angles and at different scales are extracted from the first to the twelfth layers of the network.After normalizing them into images of the same size,the higher-level features are extracted to realize the accurate recognition of multispectral image buildings.Experimental results show that compared with the traditional FCN method,this method can improve the recognition accuracy and recall rate.

Key words: multispectral image, building recognition, Full Convolution Neural network(FCN), multiscale information, training set expansion

中图分类号: