基于竞争性协同表示的局部判别投影特征提取  

Competitive collaborative representation-based local discriminant projection for feature extraction

作　　者：李静[1] 陈秀宏[1] LI Jing;CHEN Xiuhong(School of Digital Media,Jiangnan University,Wuxi 214000,China)

机构地区：[1]江南大学数字媒体学院

出　　处：《智能系统学报》2019年第5期974-981,共8页CAAI Transactions on Intelligent Systems

基　　金：江苏省研究生科研创新计划项目(KYCX17_1500)

摘　　要：特征提取算法中利用样本间的协同表示关系构造邻接图只考虑所有训练样本的协同能力,而忽视了每一类训练样本的内在竞争能力。为此,本文提出一种基于竞争性协同表示的局部判别投影特征提取算法(competitivecollaborative repesentation-based local discrininant projection for feature extraction,CCRLDP),该算法利用基于具有竞争性协同表示的方法构造类间图和类内图,考虑到邻接图中各类型系数的影响,引入保留正表示系数的思想稀疏化邻接图,通过计算类内散度矩阵和类间散度矩阵来刻画图像的局部结构并得其最优投影矩阵。在一些数据集上的实验结果表明,相比同类基于局部判别投影的特征提取算法,该算法具有很高的识别率,并在噪声和遮挡上具有良好的鲁棒性,该算法能有效地提高图像的识别效率。The feature extraction algorithm uses the cooperative representation relation between samples to construct the adjacency graph,which only considers the synergy of all training samples and ignores the competitiveness of each type of training sample.Therefore,based on competitive cooperative representation,this study proposes a local discriminant projection feature extraction algorithm and further constructs between-class and within-class graphs.Considering the influence of each type of coefficient in the adjacency graph,we introduce the idea of retaining the positive representation coefficient in the sparse adjacency graph.The local structure of the image is characterized by calculating the within-class and between-class scatter matrices;furthermore,the optimal projection matrix is obtained.The experimental results of some data sets show that compared with similar feature extraction algorithms based on local discriminant projection,the algorithm exhibits good recognition effect and good robustness in noise and occlusion and effectively increases the image recognition efficiency.

关 键 词：特征提取 协同表示 模式识别 正系数 竞争性 鲁棒性 局部结构 人脸图像 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]