计算机应用 ›› 2016, Vol. 36 ›› Issue (12): 3436-3441.DOI: 10.11772/j.issn.1001-9081.2016.12.3436

• 虚拟现实与数字媒体 • 上一篇    下一篇

全卷积网络结合改进的条件随机场循环神经网络用于SAR图像场景分类

汤浩, 何楚   

  1. 武汉大学 电子信息学院, 武汉 430072
  • 收稿日期:2016-06-01 修回日期:2016-07-22 出版日期:2016-12-10 发布日期:2016-12-08
  • 通讯作者: 汤浩
  • 作者简介:汤浩(1992-),男,湖北武汉人,硕士研究生,主要研究方向:神经网络、图像处理;何楚(1974-),男,湖南长沙人,教授,博士生导师,博士,主要研究方向:图像处理与分析、图像压缩、基于内容的图像检索、机器视觉、医学图像归档与通信系统、合成孔径雷达图像解译。
  • 基金资助:
    国家自然科学基金资助项目(61331016,41371342)。

SAR image scene classification with fully convolutional network and modified conditional random field-recurrent neural network

TANG Hao, HE Chu   

  1. School of Electronic Information, Wuhan University, Wuhan Hubei 430072, China
  • Received:2016-06-01 Revised:2016-07-22 Online:2016-12-10 Published:2016-12-08
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61331016,41371342).

摘要: 传统合成孔径雷达(SAR)图像基于粗分割像素块提取相关特征,后接支持向量机(SVM)和马尔可夫随机场(MRF)或条件随机场(CRF)进行分类,该方法存在同一像素块内部不同类别像素的误差,而且只考虑邻近区域未充分用到全局信息和结构信息。故考虑基于像素点引入全卷积网络(FCN),以ESAR卫星图像为样本,基于像素点级别构建卷积网络进行训练,得到各像素的初始类别分类概率。为了考虑全局像素类别的影响后接CRF-循环神经网络(CRF-RNN),利用FCN得到的初始概率,结合CRF结构得到全局像素类别转移结果,之后进行RNN的迭代进一步优化实验结果。由于基于像素点和考虑了全局信息与结构信息,克服了传统分类的部分缺点,使正确率较传统SVM或CRF方法平均提高了约6.5个百分点。由于CRF-RNN的距离权重是用高斯核人为拟合的,不能随实际训练样本来改变和确定,故存在一定误差,针对该问题提出可训练的全图距离权重卷积网络来改进CRF-RNN,最终实验结果表明改进后方法的正确率较未改进的CRF-RNN又提高了1.04个百分点。

关键词: 全卷积网络, 条件随机场-循环神经网络, 全局信息, 全图距离权重

Abstract: The Synthetic Aperture Radar (SAR) image uses Support Vector Machine (SVM) and Markov Random Field (MRF) or Conditional Random Field (CRF) to classify based on feature extraction of coarsely segmented pixel blocks. The traditional method exists the deviation issue of different type pixels inside the same pixel block and it only considers the adjacent area without using global information and structure information. Fully Convolutional Network (FCN) was introduced to solve the deviation problem, and the original classification probability of pixel was gotten by constructing convolutional layers based on pixel level for sample training and using ESAR images as samples. Then CRF-Recurrent Neural Network (CRF-RNN) was introduced as post layer to combine the original classification probability obtained by FCN with full image information transfer and structure information, which was produced by CRF structure. Finally, the RNN iteration was used to further optimize the experimental results. By taking advantages of global information and structure information, the proposed method based on pixel level solved some disadvantages of the traditional classification. The classification accuracy rate of the proposed method was improved by average 6.5 percentage points compared with SVM or CRF. The distance weight of CRF-RNN is fitted by Gaussian kernel, which can not be changed or determined according to the training data, thus it remains some deviation. So a convolutional network based on trainable full image distance weight was proposed to improve CRF-RNN. The experiment results show that the classification accuracy rate of the improved CRF-RNN is further improved by 1.04 percentage points.

Key words: Fully Convolutional Network (FCN), Conditional Random Field-Recurrent Neural Network (CRF-RNN), global information, full image distance weight

中图分类号: