利用贝叶斯网络融合空间上下文的高分辨遥感图像分类  

Integrating Contexts into the Classification of High-Resolution Remote Sensing Images Using the Bayesian Networks

作　　者：程环环 王润生 

机构地区：[1]ATR国防科技重点实验室,湖南长沙410073

出　　处：《计算机工程与科学》2011年第1期70-76,共7页Computer Engineering & Science

基　　金：国家自然科学基金资助项目(60373000)

摘　　要：针对高分辨遥感图像,本文提出了一种基于贝叶斯网络的上下文模型,以及基于该模型的面向对象的遥感图像分类方法。首先,利用支持向量机(SVM)实现分割区域的初始分类,获得各个类别的候选区域。然后,利用提出的上下文模型融合候选区域及其周围区域的上下文信息,通过贝叶斯网络推理,将候选区域分类到各类地物类型中。基于贝叶斯网络的上下文模型由候选区域节点、相关区域节点和上下文节点三部分组成。对于不同类型的地物,通过贝叶斯网络的结构学习算法学习得到不同的空间关系作为上下文节点。因此,该模型能够针对不同的地物类别利用不同的空间上下文信息,使得分类过程更智能和有效。实验结果表明,本文提出的算法能够很好地利用上下文信息,对高分辨遥感图像中的各种地物进行有效的分类和检测。In this paper, a Bayesian networkbased context model （RCBN） is presented to classify high resolution remote sensing （HRRS） images. First of all, image regions are classified by SVMs and candidate regions for each ground cover types are obtained. Then, the hybrid streams of candidate regions and spatial context information are piped into the context model, which will produce the category labels of regions by performing inference through the network. The RCBN consists of three kinds of nodes, which are nodes for candidate regions, related regions and contexts. The context nodes vary with different ground cover types, which are learned by the structure learning algorithm from training samples. Therefore, our RCBN model is capable of using the specific context information for each ground cover type, which makes the classification process more intelligent and efficient. The performance of the approach is evaluated qualitatively and quantitatively with comparative experiments, and the results show that the proposed methods are able to automatically classify and detect segments belonging to the same object classes.

关 键 词：高分率遥感 图像分类 上下文信息 贝叶斯网络 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]