基于关联分层条件随机场的高分辨率影像分类方法研究

项目名称： 基于关联分层条件随机场的高分辨率影像分类方法研究

项目编号： No.41301386

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 天文学、地球科学

项目作者： 杨耘

作者单位： 长安大学

项目金额： 25万元

中文摘要： 甚高空间分辨率(VHR)遥感影像高精度自动分类这一技术难题制约了其应用水平，而融合多尺度的目标空间及语义信息、场景先验是有效解决这一难题的关键技术。条件随机场(CRFs)在表达远距离信息交互及概率建模方面有其优势,而分层CRFs还能表达目标多层次结构及语义信息。针对城区VHR 影像分类中存在的难题,本项目利用CRFs优势，拟开展基于"对象-目标-场景"逐层关联的分类方法研究。创新性研究:(1)基于分层影像解析的思想，拟建立一个逐层关联的分层CRFs分类模型。该分类器可集成目标空间上下文和多重语义关系，以缓解分割错误对分类的不利影响;(2)通过挖掘潜在的场景知识，拟研究考虑相邻对象类标签属性的交互势函数定义方法;此外,针对模型推理问题,拟基于语义分析的思想实现对重参数化的图割推理方法的优化。最终目标是基于关联分层CRFs模型框架，寻求一种基于影像理解和认知的VHR影像分类方法。

中文关键词： 关联分层条件随机场；分割质量；高分辨率遥感；影像分类；多尺度分割

英文摘要： It is a difficulty to automatically classify and interpret remotely sensed imagery with very high spatial resoultion(VHR) ensuring high precision and automatication, so as to prevent an extensive application of remotely sensed imagery. To solve the difficulity, it is a key to integrate spatial information from different scales,and semantic information from a multiscale image analysis as well as scene prior etc..Conditional random fields(CRFs) have the advantages of interacting over a long spatial range each other and conveniently modelling posteri-probability, More important,hierarchical CRFs is capable of expressing multi-scale objects feature and semanic information better.For the difficulties in VHR remote sensing image classification esp.in urban city,the project will develop a classification method with "segments-objects-scene" associative process layer by layer and the advantages over CRFs theory. The innovative studies are: (1)For the problem of how to integrate mutl-scale spatial,semanic information and, we will develop an associative hierarchical CRFs for VHR remote sensing image classification from the view of multi-layer scene analysis. (2)To alleviate the misclassification esp.,betweeen buildings and roads without the help of elevation data,latent scene knowledge is mined and used to extend traditio

英文关键词： Associative hierarchical Conditional Random Field；segmentation quality；high resolution remote sensing；image classification；multiscale segmentation