计算机工程
計算機工程
계산궤공정
COMPUTER ENGINEERING
2014年
9期
243-247
,共5页
线性鉴别分析%直接线性鉴别分析%二维主成分分析%小样本问题%人脸识别%特征提取
線性鑒彆分析%直接線性鑒彆分析%二維主成分分析%小樣本問題%人臉識彆%特徵提取
선성감별분석%직접선성감별분석%이유주성분분석%소양본문제%인검식별%특정제취
Linear Discriminant Analysis ( LDA )%Direct LDA ( DLDA )%Two Dimension Principle Component Analysis(2D-PCA)%small sample size problem%face recognition%feature extraction
线性鉴别分析( LDA)小样本问题的已有解决方法在构造最优投影子空间时未完整利用LDA的4个信息空间,为此,提出一种基于二维主成分分析(2D-PCA)的两级LDA人脸识别方法。采用减法运算对样本类内散度矩阵和类间散度矩阵的特征值矩阵求逆,以解决小样本问题,并连续应用Fisher准则和修改后的Fisher准则连接2个投影子空间,获取包含LDA的4个信息空间的最优投影方向,利用2 D-PCA对输入样本做预处理,以减少计算复杂度。在ORL和YALE人脸库上的实验结果表明,该方法虽然训练时间略有增加,但识别率分别为92.5%和95.8%,优于其他常用LDA算法。
線性鑒彆分析( LDA)小樣本問題的已有解決方法在構造最優投影子空間時未完整利用LDA的4箇信息空間,為此,提齣一種基于二維主成分分析(2D-PCA)的兩級LDA人臉識彆方法。採用減法運算對樣本類內散度矩陣和類間散度矩陣的特徵值矩陣求逆,以解決小樣本問題,併連續應用Fisher準則和脩改後的Fisher準則連接2箇投影子空間,穫取包含LDA的4箇信息空間的最優投影方嚮,利用2 D-PCA對輸入樣本做預處理,以減少計算複雜度。在ORL和YALE人臉庫上的實驗結果錶明,該方法雖然訓練時間略有增加,但識彆率分彆為92.5%和95.8%,優于其他常用LDA算法。
선성감별분석( LDA)소양본문제적이유해결방법재구조최우투영자공간시미완정이용LDA적4개신식공간,위차,제출일충기우이유주성분분석(2D-PCA)적량급LDA인검식별방법。채용감법운산대양본류내산도구진화류간산도구진적특정치구진구역,이해결소양본문제,병련속응용Fisher준칙화수개후적Fisher준칙련접2개투영자공간,획취포함LDA적4개신식공간적최우투영방향,이용2 D-PCA대수입양본주예처리,이감소계산복잡도。재ORL화YALE인검고상적실험결과표명,해방법수연훈련시간략유증가,단식별솔분별위92.5%화95.8%,우우기타상용LDA산법。
Aiming at the existing algorithms which do not use the whole four information space of Linear Discriminant Analysis(LDA) in solving the small sample size problem,a two-stage LDA face recognition algorithm based on Two Dimension Principle Component Analyses ( 2D-PCA ) is proposed. The small sample size problem is solved by a subtraction to estimate the inverse matrix of the eigenvalues matrix of the singular with-class scatter matrix and between-class scatter matrix. Thus,the projection subspaces resulting from continuously using the traditional Fisher criterion and a modified Fisher criterion,are concatenated to obtain the optimal projection space including whole four information space of LDA. To reduce the computational complexity,the 2D-PCA is used to preprocess on input samples. The recognize rates of the proposed algorithm on ORL and YALE database are 92 . 5% and 95 . 8% which are higher than other LDA algorithms despite the slightly increase of training time.