图像多分类主动学习方法  

Multi-class active learning approach for image classification

作　　者：刘君[1,2] 熊忠阳[1] 王银辉[1] 

机构地区：[1]重庆大学计算机学院,重庆400030 [2]重庆广播电视大学理工学院,重庆400052

出　　处：《计算机工程与应用》2010年第13期11-14,27,共5页Computer Engineering and Applications

基　　金：中国博士后科学基金资助项目(No.20070420711);重庆市科委自然科学基金计划资助项目(No.2007BB2372)

摘　　要：以决策速度快的决策导向非循环图支持向量机(Decision DirectedAcyclic Graph Support Vector Machine)为基准分类器,结合主动学习的思想,提出了一种图像多分类主动学习方法。这种方法是一种半自动的图像语义分类方法,可以将图像分成多个语义类别。该方法在最近边界主动选择方法的基础上,提出一种基于质疑度的主动选择策略。这种策略将SVMactive中提出的最近邻SVM分类面选择的反馈样例策略延伸到多分类中,通过区别对待奇异样例和容易错分样例,减少了噪声数据对分类器的干扰,提高了分类的精度。A multi-class active learning approach for image classification is described by using Decision Directed Acyclic Graph Support Vector Machine,featured by rapid decision-making,as basic classifier and the ideas of active learning are combined.This approach is a semi-automatic classification method for image semantics,and can be used to divide images into different categories.In this approach,an active selection strategy based on the level of doubts is presented on the basis of the widely-used method of selecting the closest samples to the dividing hyperplane.This strategy extends the strategy of selecting the closest ones to the SVM dividing hyperplane specified in SVMactive algorithm to multi-class classification,and at the same time,treats odd samples and easily mistaken samples differently.In this way,the interference of noise on classifier is reduced,and the accuracy of classification is increased.

关 键 词：支持向量机 多分类 决策导向非循环图 主动学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]