Eigen-space Method based on Class-features (EMC), a variant of Multiple Discriminant Analysis (MDA), has been proposed and applied for automatic facial expression recognition. Although EMC was reported to outperform MDA in Ref. [1][2], no mathematical explanations for the difference of performance have been given. In the present paper, we will first reformulate MDA and EMC based on a new model of Maximum Log Likelihood (MLL) estimation. By using this model, we will explain from the perspective of statistical inference that the difference of the underlying mechanism locates in that EMC is a variant of MDA with lower degree of freedom by assuming the covariance to be sphered in all directions. A thorough comparison between EMC and MDA in robust recognition of facial expressions will also be made to verify our conclusion that EMC outperforms MDA because it is more robust against over-fitting due to its lower degree of freedom.
展开▼