基于特征距离的多类SVM分类方法研究

Research on Multi-class SVM Classification Method Based on Jeffries-Matusita Distance

提出了一种基于特征分离性测度的面向对象分类方法。首先利用区域增长分割影像获得影像对象,并计算光谱、纹理、形状等多种分类特征,然后在构建多类SVM分类器过程中,对于任意两个分类类别对,利用Jeffries-Matusita距离选择最合适的特征。实验证明,相比于原始方法,基于Jeffries-Matusita距离的多类分类器能够有效减少建筑物、道路等复杂地物的误分现象,提高分类的总体精度和Kappa系数。

This paper presented an object-oriented classification method based on separability measurement. Image objects were obtained by region growing segmentation, and many different kinds of characteristics were calculated for the image objects, such as spectral, texture and shape at first. And then, a new multi-class SVM classifier was constructed in the one-against-one way, and the most suitable characteristic set were selected for every two-class-pair by JeffriesMatusita distance. The experiment results show that the new multi-class SVM classifier based on Jeffries-Matusita distance can reduce wrong classification for complicated feature, such as building and road, and improve total accuracy and Kappa coefficient significantly.