Abstract: To the question that extracted dimension of feature coefficient is great based on the traditional two-dimensional linear discriminant analysis，This paper proposes an Improved Bilateral Two-Dimensional Linear Discriminant Analysis（GB2DLDA）.The within-class scatter matrix and the between-class scatter matrix are compressed from two directions，two Fisher discriminant function are formed with the compressed scatter matrix，and then，face image is projected to the projection matrixes，the feature coefficient is extracted.Experiment shows that the dimension of feature coefficient is less than other methods on the same classification accuracy.Choosing suitable feature vector，classification accuracy is better than other methods.

Key words: Two-Dimensional Principal Component Analysis（2DPCA）, Bilateral Two-Dimensional Linear Discriminant Analysis（B2DLDA）, Improved Bilateral Two-Dimensional Linear Discriminant Analysis（GB2DLDA）, compress, projection matrix

摘要： 针对传统的二维线性判别方法提取出的人脸特征系数维数大的问题，提出一个改进的双向二维线性判别分析方法GB2DLDA。双向压缩类内和类间散布矩阵，用压缩后的散布矩阵构成两个Fisher鉴别准则函数，求出两个投影矩阵，然后人脸图像矩阵向投影矩阵投影，提取出特征系数。实验证明在相同识别率下，用此方法提取的特征系数维数明显少于其它二维线性判别分析方法。在选择合适的特征向量的情况下，此方法的识别率要好于其它二维线性判别分析方法。

关键词: 二维主元分析法, 双向二维线性鉴别分析方法, 改进的双向二维线性判别分析方法, 压缩, 投影矩阵