基于ELM和证据理论的纹理图像分类  

Texture Image Classification Based on ELM and Evidence

作　　者：易丐[1] 李国进[1] 王祥铜 YI Gai LI Guo-jin WANG Xiang-tong(College of Electrical Engineering, Guangxi University, Nanning,Guangxi 530004, Chin)

机构地区：[1]广西大学电气工程学院,广西南宁530004

出　　处：《计算技术与自动化》2017年第1期98-102,共5页Computing Technology and Automation

摘　　要：目前纹理图像分类有不同的方法,但对纹理的描述还不够全面,而且当有新方法提取的特征加入时,系统的可扩展性也不够,通用性不好。本文针对上述问题提出了一种将D-S证据理论与极限学习机相结合的决策级融合模型,用来对纹理图像进行分类。采用三种不同方法来提取特征以获得更多更全面的纹理表现形式,并对提取的每种特征向量用极限学习机建立相应的分类器,最后用D-S证据理论在不确定性表示、度量和组合方面有着的优势来进行决策级融合。对于证据理论中基本概率赋值函数(BPAF)难以有效获取的问题,由于极限学习机具有学习速度快,泛化性能好的优点并且产生唯一的最优解的优点,所以利用其来构造其基本概率赋值函数。实验结果表明这种方法比单个分类器具有更高的识别正确率,降低了识别的不确定性。There are different methods for texture image classification at present. In this paper, we propose a multi clas sifter decision level fusion model based on D-S evidence theory and extreme learning machine, which is used to classify the texture image, we use three different methods to extract texturefeature of texture imageto obtain more more comprehensive texture forms and build classifiers using extreme learning machine based on the each feature vector. Finally the classifiers were fused by D-S evidence theory since the advantage of uncertainty representation, measure and the aspect of combination. To the problem of getting the basic probability assignment function （BPAF） in I〉S evidence theory is hard ,we used ELM to construct the basic probability assignment function, because the extreme learning machine have advantages of fast learning speed, good generalization performance and produce the uniqueness of the optimalsolution. The experimental results show that this method has higher recognition accuracy than a single classifier, and the uncertainty of the recognition is reduced.

关 键 词：纹理图像分类 特征提取 D-S证据理论 极限学习机 基本概率赋值函数 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]