基于局部类内结构的鉴别性字典学习方法

针对支持向量引导的字典学习(support vector guided dictionary learning,SVGDL)的鉴别约束项只体现了大间隔原理,而没有很好地利用数据空间内在结构信息的问题,提出了一种新颖的鉴别性字典学习方法——基于局部类内结构的鉴别性字典学习方法。该方法结合了大间隔原理和局部Fisher线性鉴别分析作为鉴别约束条件来指导指点学习。通过建立一个局部类内散度矩阵来编码数据空间的分布结构,增强了挖掘同类数据空间局部结构的能力并进一步地表示了编码向量在数据空间中的局部相似性。为了评价提出方法在图像识别上的表现,在几个常见图像数据集上进行了实验。结果表明,提出方法与大间隔方法相比,在平均识别率上有着明显的提高。

Discriminative dictionary learning based on locality intra-class structure

Aiming at the limitation that the discriminative term of only embodies the max-margin principle as SVM,and fails to utilize the intrinsic structure information of data space,this paper proposed a novel discriminative dictionary learning method called discriminative dictionary learning based on locality intra-class structure(LCSDDL).The proposed method combined max-margin principle with local Fisher discriminant analysis(LFDA)as the discriminative term to guide dictionary learning.This method constructed a local within-class scatter matrix to encode the local structure of data space,which enhanced the ability of exploiting the local structure of same class data space and further reflected the local similarity of coding vectors in data space.In order to evaluate the performance of the proposed method for image recognition,the experiment carried on several common datasets.From the experimental result,the proposed method has an obvious improvement over the other competing methods.