计算机工程

• 图形图像处理 • 上一篇    下一篇

基于改进全卷积神经网络的航拍图像语义分类方法

易盟,隋立春   

  1. (长安大学 电子与控制工程学院,西安 710064)
  • 收稿日期:2017-01-25 出版日期:2017-10-15 发布日期:2017-10-15
  • 作者简介:易盟(1982—),男,讲师、博士,主研方向为机器学习、计算机视觉、虚拟现实;隋立春,教授、博士。
  • 基金项目:
    中国博士后科学基金(2016M590912);中央高校基本科研业务费专项资金(310832151097)。

Aerial Image Semantic Classification Method Based on Improved Full Convolution Neural Network

YI Meng,SUI Lichun   

  1. (Institute Electronic and Control Engineering,Chang’an University,Xi’an 710064,China)
  • Received:2017-01-25 Online:2017-10-15 Published:2017-10-15

摘要: 现有的卷积神经网络方法难以对图像的每个像素进行语义识别,较难从像素层面分解出图像的不同类别。为此,提出一种端到端的全卷积深度网络,以实现高分辨航拍图像像素级的语义分割及识别。通过全卷积神经网络对图像强度信息和地理信息系统信息分别采用独立通道进行处理,在全卷积神经网络的最终层合并2个通道,并对每个像素进行全连接像素级标注,利用条件随机场作为后期处理方法平滑相似区域,同时保留图像中的边缘信息。实验结果表明,与传统视觉语义分类算法相比,该算法在航拍图像像素级分类上的准确率更高,识别效果更好。

关键词: 图像分类, 语义标注, 神经网络, 目标检测, 条件随机场

Abstract: The existing Convolution Neural Networks(CNNs) method cannot semantically identify each pixel,and it is difficult to decompose the different types of images from the pixel level.Therefore,an end-to-end full-convolution depth network is proposed to achieve high-resolution aerial image pixel level semantic segmentation and recognition.Full convolution neural network is used to process the image intensity information and Geographical Information System(GIS) information with independent channel,two channel results are merged at the final layer of full convolution neural network,and each pixel is labeled at fully connected pixel level.The Conditional Random Field(CRF) is used as the post-processing method to smooth the similar region,while preserving the edge information in the image.Experimental results show that the proposed algorithm has higher accuracy and better recognition rate than the traditional visual semantic classification algorithm.

Key words: image classification, semantic annotation, neural network, target detection, Conditional Random Field(CRF)

中图分类号: