现代电子技术
現代電子技術
현대전자기술
MODERN ELECTRONICS TECHNIQUE
2013年
22期
24-27
,共4页
翟佳%谭龙%潘静%庞彦伟
翟佳%譚龍%潘靜%龐彥偉
적가%담룡%반정%방언위
线性判别分析%无穷范数%二值化%特征提取
線性判彆分析%無窮範數%二值化%特徵提取
선성판별분석%무궁범수%이치화%특정제취
linear discriminant analysis%infinite norm%binarization%feature extraction
线性判别分析(LDA)是监督式的特征提取方法,在人脸识别等领域得到了广泛应用。为了提高特征提取速度,提出了基于无穷范数的线性判别分析方法。传统LDA方法将目标函数表示为类内散布矩阵和类间散布矩阵之差的或者商的L2范数,且通常需要涉及到矩阵求逆和特征值分解问题。与传统方法不同,这里所提方法将目标函数表示为类内散布矩阵和类间散布矩阵之差的无穷范数,而且最优解是以迭代形式得到,避免了耗时的特征值分解。无穷范数使得到的基向量实现了二值化,即元素仅在-1和1两个数字内取值,避免了特征提取时的浮点型点积运算,从而降低了测试时间,提高了效率。在ORL人脸数据库和Yale数据库上的实验表明所提算法是有效的。
線性判彆分析(LDA)是鑑督式的特徵提取方法,在人臉識彆等領域得到瞭廣汎應用。為瞭提高特徵提取速度,提齣瞭基于無窮範數的線性判彆分析方法。傳統LDA方法將目標函數錶示為類內散佈矩陣和類間散佈矩陣之差的或者商的L2範數,且通常需要涉及到矩陣求逆和特徵值分解問題。與傳統方法不同,這裏所提方法將目標函數錶示為類內散佈矩陣和類間散佈矩陣之差的無窮範數,而且最優解是以迭代形式得到,避免瞭耗時的特徵值分解。無窮範數使得到的基嚮量實現瞭二值化,即元素僅在-1和1兩箇數字內取值,避免瞭特徵提取時的浮點型點積運算,從而降低瞭測試時間,提高瞭效率。在ORL人臉數據庫和Yale數據庫上的實驗錶明所提算法是有效的。
선성판별분석(LDA)시감독식적특정제취방법,재인검식별등영역득도료엄범응용。위료제고특정제취속도,제출료기우무궁범수적선성판별분석방법。전통LDA방법장목표함수표시위류내산포구진화류간산포구진지차적혹자상적L2범수,차통상수요섭급도구진구역화특정치분해문제。여전통방법불동,저리소제방법장목표함수표시위류내산포구진화류간산포구진지차적무궁범수,이차최우해시이질대형식득도,피면료모시적특정치분해。무궁범수사득도적기향량실현료이치화,즉원소부재-1화1량개수자내취치,피면료특정제취시적부점형점적운산,종이강저료측시시간,제고료효솔。재ORL인검수거고화Yale수거고상적실험표명소제산법시유효적。
The linear discriminant analysis(LDA)is a method of supervised feature extraction. It has been widely used in the field of computer vision such as face recognition. An infinite norm based LDA method is proposed this paper to improve the efficiency of feature extraction. Traditional LDA methods express their objective functions as either difference of between-class scattering matrix and within-class scattering matrix or quotient in the L2 norm. Consequently,these methods need to involve in matrix inversion and eigen-value decomposition. By contrast,the proposed method utilizes L-norm(infinite norm)instead of L2 norm to formulate the objective function with respect to the difference between between-class scatter matrix and within-class scat-ter matrix. Because the solution is obtained iteratively,this method avoids time-consuming eigen-decomposition. Moreover,the projection vector realizes binarization,and the value of elements is -1 or 1,resulting in high efficiency because it avoids compu-ting the inner product between a sample and the projection vector. The results of experiments in ORL database and Yale data-base demonstrate the efficiency and effectiveness of the proposed method.