最小类内方差支持向量引导的字典学习

Dictionary Learning Guided by Minimum Class Variance Support Vector

支持向量引导的字典学习算法依据大间隔分类原则,仅考虑每类编码向量边界条件建立决策超平面,未利用数据的分布信息,在一定程度上限制了模型的泛化能力.为解决该问题,提出最小类内方差支持向量引导的字典学习算法.将融合Fisher线性鉴别分析和支持向量机大间隔分类准则的最小类内方差支持向量机作为鉴别条件,在模型分类器的交替优化过程中,充分考虑编码向量的分布信息,保障同类编码向量总体一致的同时降低向量间的耦合度并修正分类矢量,从而挖掘编码向量鉴别信息,使其更好地引导字典学习以提高算法分类性能.在人脸、物体和手写数字识别数据集上的实验结果表明,在大部分样本和原子数量条件下,该算法的识别率和原子鲁棒性均优于K奇异值分解、局部特征和类标嵌入约束等经典字典学习算法.

Existing Support Vector Guided Dictionary Learning(SVGDL)algorithm based on the principle of large-margin classification.When establishing decision-making hyperplanes,the algorithms consider only the boundary conditions of each class of encoding vectors,but ignore data distribution information,which limits the generalization ability of the model.To address the problem,this paper proposes a Minimum Class Variance Support Vector Guided Dictionary Learning(MCVGDL)algorithm.First,MCVGDL takes the Minimum Class Variance Support Vector Machine(MCVSVM),which combines Fisher linear discriminant analysis and the large margin classification principle of Support Vector Machine(SVM),as discriminant term.Second,during alternate optimization of model classifiers,MCVGDL comprehensively takes the distribution information of encoding vectors into account,to guarantee the overall consistency of encoding vectors of similar samples and reduce the coupling degree of corresponding components between vectors and modifies SVM classification vectors.So,the discriminant information of encoding vectors can be fully mined to better guide dictionary learning,improving the classification performance.Experimental results on face,object,and handwritten digit recognition datasets show that in terms of the recognition rate and atomic robustness,the proposed algorithm outperforms classical dictionary learning algorithms,including K Singular Value Decomposition(KSVD)and Local Constrained and Label Embedding Dictionary Learning(LCLE-DL),etc.