基于2维偏最小二乘法的图像局部特征提取及其在面部表情识别中的应用

Image Local Feature Extraction Method Based on Two-dimensional Partial Least Square and Its Application in Facial Expression Recognition

为了更有效地提取图像的局部特征,提出了一种基于2维偏最小二乘法(two-dimensional partial leastsquare,2DPLS)的图像局部特征提取方法,并将其应用于面部表情识别中。该方法首先利用局部二元模式(localbinary pattern,LBP)算子提取一幅图像中所有子块的纹理特征,并将其组合成局部纹理特征矩阵。由于样本图像被转化为局部纹理特征矩阵,因此可将传统PLS方法推广为2DPLS方法,用来提取其中的判别信息。2DPLS方法通过对类成员关系矩阵的构造进行相应的修改,使其适应样本的矩阵形式,并能体现出人脸局部信息重要性的差异。同时,对于类成员关系协方差矩阵的奇异性问题,也推导出了其广义逆的解析解。基于JAFFE人脸表情库的实验结果表明,该方法不但可以有效地提取图像局部特征,并能取得良好的表情识别效果。

In this paper, we proposed a local feature extraction method based on a two-dimensional partial least square （2DPLS） for images, and then applied it in the facial expression recognition. Firstly, this method combines texture features of all sub windows of an image extracted by local binary pattern（LBP） into a local texture feature matrix. In order to extract the discrimination information, the traditional PLS method is extended to 2DPLS method since the images have been transformed to local texture matrices. In the 2DPLS method, the class-membership matrix is modified to adapt to the matrix form of the samples and represents the difference of the importance of the local image information. Meanwhile, the analytic form of the generalized inverse of class-membership matrix is derived. The experiment results based on JAFFE database show that the proposed method can effectively extract the local feature from images and achieve good performance in facial expression recognition.