【摘要】 特征提取是人脸识别的关键。特征提取方法一般需要预先把二维图像转化成一维图像向量。然而高维的图像向量会导致不能快速、精确地计算所需的协方差矩阵及其特征向量。针对该问题,提出了一种基于二维局部鉴别高斯的特征提取方法(2D-LDG)。该方法继承一维局部鉴别高斯降维方法的优点,其目标函数是留一交叉验证误差的光滑逼近,并且只考虑训练样本的局部分布,对训练样本的全局分布不做任何假设。同时,2D-LDG直接对二维图像做特征提取,不需要事先把图像转化为维数巨大的图像向量,能快速、精确地计算协方差矩阵及其特征向量。在ORL、YaleB人脸数据库上的实验结果表明,2D-LDG特征提取方法有良好的识别效果。更多还原

【Abstract】 Feature extraction plays an important role in face recognition.In general,feature extraction methods need to transfer the 2Dimages into 1Dvectors.As a result,it is hard to calculate the covariant matrix and eigen-vector efficiently and exactly due to the high dimensionality.This paper proposed a new feature extraction method named 2Dlocal discriminant Gaussian(2D-LDG).It inherits the properties of LDG and the objective function of proposed method is also an approximation to the leave-one-out training error of a local quadratic discriminant analysis classifier.Also,it applies local Gaussians to eastimate probability in each point,relaxing the assumption on the class probability density function. Meanwhile,2D-LDG is operated on 2Dimages directly which avoids turning the image matrixes into high dimensional vectors,and is able to calculate the covariance matrix and eigen-vector in a more efficient and accurate way.Experiments on ORL and YaleB-Extended show that our proposed 2D-LDG feature extraction method achieves better performance in face recognition.更多还原