Abstract: A Linear Discriminant Analysis method Base on Column（CBLDA） is proposed. CBLDA calculates the projection matrix for each class. Considered the strong symmetry of face images, the selection of the nearest of the column for projection matrix will conquer some variations of illuminations and postures. So the projection matrix should be maximized the between-class nearest columns and minimized the within-class nearest columns. Also, columns are the inner scale of the face image, which will be changed according to the face image resolution. It does not need to decide the size of image block. Experimental results on ORL, FERET and YALE B face databases show that the proposed method is more robust than several state-of-the-art face recognition methods, 2DLDA, 2DLPP and 2DLGEDA.

Key words: Linear Discriminant Analysis（LDA）, face recognition, local method, 2D linear discriminant analysis method

摘要： 人脸识别是模式识别中重要的研究内容，具有广泛的应用前景。为了进一步提高人脸识别中线性鉴别方法的鲁棒性，提出了一种基于列最近邻的线性鉴别方法（CBLDA）。CBLDA为每一类找到一个投影矩阵，使得人脸图像中的每一列经过投影矩阵投影后，能够更靠近类内列最近邻同时离类间列最近邻越远。当测试样本与经过其类别的投影矩阵投影后能够得到更有利于分类的结果。CBLDA类似于分块或者子图的方法，选择最近邻列作为分块的策略的主要优点：（1）列是图像的固有尺寸，会随分辨率的变化而变化，因此不需要决定分块的大小；（2）人脸具有对称性，对列求得类内列最近邻可以较好克服一些左右姿态和光照变化的影响，提高算法的鲁棒性。为了验证CBLDA的有效性，在ORL和FERET人脸数据库中与2D-LDA、2D-LPP和2D-LGEDA等二维算法进行了对比实验，结果表明CBLDA在识别率有大幅的提升，证明了算法的有效性。

关键词: 线性鉴别分析, 人脸识别, 局部方法, 二维线性鉴别分析