This paper suggests a Dual Stage MLP, a new ensemble structure for facial expression recognition, and Subset Pre-Training which supports the structure as an effective learning method. Subset Pre-Training is the asymmetrical pre-training method in which 1) some parts of training data are trained to MLPs in the first stage, 2) then the whole training data are utilized to train MLP in the second stage. Facial expression recognition was carried out based on Extended Cohn Kanade database to demonstrate the effectiveness of the proposing method, and the Local Binary Pattern was set as the feature for an experiment. The result of the experiment confirmed the bigger asymmetry guarantees the higher classification rates. Compared with the entire time needed in learning the whole training data, the proposing method reduced the time by one eighth while showing the improved classification accuracy rate 99.25%.
展开▼
