摘要
鉴于线性判别分析(Linear Discriminant Analysis,LDA)算法存在的弊端,本文提出了一种基于局部离散度的监督型线性判别分析(Supervised Linear Discriminant Analysis based on Local Dispersion,SLDALD)算法.新方法的改进主要有:1)从像元邻域的角度出发,对类内散布矩阵、类间散布矩阵进行重新定义,得到类内邻域散布矩阵和类间邻域散布矩阵.新定义充分考虑了不同区域之间像元光谱特征离散度的差异性;2)在计算类间邻域散布矩阵时,赋予类边界像元较大的权重,让特征降维更针对此类像元;3)在计算类内邻域散布矩阵时,加大类边界像元的权重,让后续的特征降维针对此类像元.同时,降低噪声点的权重,以抑制噪声点对特征降维的干扰.实验结果表明:相比依据LDA算法所获得的低维特征的分类结果,以SLDALD算法所获得的低维特征为依据,影像分类精度得到明显地提高.
In consideration of the disadvantages of Linear Discriminant Analysis(LDA)algorithm,this paper puts forward a supervised liner discriminant analysis based on local dispersion algorithm that with the following improvements:1)from the perspective of pixel neighborhood,the intra-class scatter matrix and the inter-class scatter matrix are redefined to obtain the intra-class scatter matrix and the inter-class scatter matrix.The new definition fully considers the differences of spectral dispersion in different regions.2)in the calculation of the neighborhood scattering matrix in classes,the class boundary pixel is given a larger weight,so that the feature dimensionality reduction is more targeted at such pixels;3)when calculating the neighborhood scattering matrix within the class,increase the weight of the class boundary pixel,and make subsequent feature dimensionality reduction for such pixel.At the same time,the weight of noise points is reduced to suppress the interference of noise points on feature dimensionality reduction.The experimental results show that compared with the classification results of low-dimensional features obtained by LDA algorithm,the classification accuracy of images is obviously improved by using the low-dimensional features obtained by SLDALD algorithm.
作者
孙小丹
陈文
SUN Xiaodan;CHEN Wen(Fuzhou Polytechnic,Fuzhou,Fujian 350108)
出处
《绵阳师范学院学报》
2019年第11期77-84,91,共9页
Journal of Mianyang Teachers＇ College
基金
福建省自然科学基金项目(2016J01337)
福建省教育厅科研项目(JAT171060)
关键词
特征降维
类内邻域散布矩阵
类间邻域散布矩阵
影像分类
feature dimensionality reduction
intra-class scatter matrix
inter-class scatter matrix
classification of image
