摘要： 传统的模式识别方法认为特征是相互独立的,容易忽略多模态特征之间多元的关联性,从而造成识别的误差。为此,基于超图模型,提出一种新的特征整合方法。定义共享熵的计算方法用以表示多个特征之间的关联程度,以每个特征作为顶点,特征之间的多元关系作为超边。对形成的超图,定义模块度函数取代传统的切边数,作为衡量子超图的社团特性强弱的指标,应用超图分割算法,对原始的多模态特征进行聚类划分。在划分集合上采用多分类Boosting方法,形成最终的强分类器。实验结果表明,与线性支持向量机、多核学习等当前流行的特征融合方法相比,该方法能有效提高识别准确率。

关键词: 超图, 多模态特征, 共享熵, 模块度, 分类器

Abstract: Features are usually considered as independent of each other in traditional pattern recognition methods.The neglect of the correlation among multimodal features is part of the reason for recognition error.Aiming at integrating multimodal features,this paper presents a hypergraph framework.Under hypergraph model,this paper defines a new measure called shared entropy to capture the multivariate correlation among the multimodal features.Each feature is abstracted as a vertex,and if the value of shared entropy reaches the threshold,a hyperedge can be built.Then,the hypergraph is clustered into a set of partitions using the modularity instead of cut-edges to measure the community degree of sub-hypergraphs.Finally,combining the weak classifiers learned from each partition,a multiclass Boosting method is used to form the last strong classifier.Experimental results show that this method can improve the recognition accuracy effectively compared with the current popular methods such as linear Support Vector Machine(SVM) and Multiple Kernel Learning(MKL).

Key words: hypergraph, multimodal feature, shared entropy, modularity, classifier

中图分类号: 

• TP18