Abstract:
|
Face recognition can be viewed as an image classification problem. Linear discriminant analysis is a widely used classification method which aims to find a low dimensional subspace where the groups are as separated as possible. However, for face recognition, the dimension of each image is usually larger the sample size, which makes the original linear discriminant analysis intractable. Many algorithms have been proposed to overcome this ill-posed issue. In this talk, we propose to have a rough dimension reduction step by multilinear principal component analysis following by partial inverse regression. For this proposed algorithm, we have to determine the dimension for the multilinear principal component step. We propose to check the ratio of the between group variation to the total variation. When this ratio drops, it indicates that some useful between group information is lost after dimension reduction, which naturally provides a lower bound of the dimension. Based on this ratio, we propose a complete algorithm which produces a higher classification accuracy on some benchmark image data sets.
|