航拍图像区域多特征紧耦合多级分类算法  

Multi-stage Classification Algorithm Based on Multi-feature Tightly Coupled for Aerial Image

作　　者：马旭[1] 程咏梅[1] 郝帅[1] 赵建涛[1] 王涛[2] 

机构地区：[1]西北工业大学自动化学院,陕西西安710072 [2]中国航空工业集团西安飞行自动控制研究所重点实验室,陕西西安710065

出　　处：《微电子学与计算机》2014年第10期13-17,共5页Microelectronics & Computer

基　　金：西安市科技计划项目(CXY1350(2));国家自然科学基金重点项目(61135001)

摘　　要：提出一种基于多特征紧耦合的航拍图像区域多级分类算法.首先将样本图像从RGB空间转换到HSV空间,分别提取颜色矩特征和Gabor纹理特征,并对Gabor纹理特征进行PCA降维;然后将两种特征相乘构成紧耦合矩阵,进一步生成紧耦合特征向量,并对生成的紧耦合特征向量再次进行PCA降维;接着搭建了由5个概率神经网络分类器构成的多级分类器.最后利用Google Earth软件截取不同时间、不同尺度的图像,作为训练样本和测试样本,进行多级分类器的训练和测试.实验结果表明,相比于单特征及多特征松耦合的分类方法,提出的方法分类精度较高.A multi-stage classification algorithm based on multi-feature tightly coupled for aerial image is proposed . Firstly ,the sample images are transformed from RGB space to HSV space .Color moment feature and Gabor texture feature are extracted and PCA is used to reduce the feature dimensionality of Gabor texture .Secondly ,tightly coupled matrix is gotten by multiplying the two characteristics and tightly coupled vector is generated .PCA is used again to reduce the feature dimensionality of tightly coupled vector .Then a multi-stage classifier is constructed by five probabilistic neural network classifiers .Lastly ,images of Xi’an Qing Zhen Village with different time and scale are selected by Google Earth software .Grassland ,lakes ,trees ,houses and land from the images are selected as training samples and test samples . And these samples are used for multi-stage classifier training and testing . Experimental results show that proposed method has higher classification accuracy compared with single feature classification method and the method of multi-feature loosely coupled .

关 键 词：多特征 紧耦合 航拍图像 多级分类器 概率神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]