| As the key technologies of face recognition, feature selection and classifier design are always the hot points of FR research. It is normal to improve the performance of one face recognition system if it takes an effective schedule to select feature and build the classifier. Therefore, the target of this paper is to build one whole face recognition algorithm from three aspects: feature description, feature selection and classifier design.In the first part of this paper, the frame of face recognition is described. And then the key technology such as feature description, feature selection and classifier design are introduced briefly, and it is emphasized that more work should be done on these areas. In the following, the variance of illumination, pose and the dodge, which are the key problems of face recognition, are introduced simply. Based on these problems in face recognition area, local feature is taken to describe the faces for its robust to illumination variance, pose variance and dodge.Then, the flow chart of the method in this paper is proposed. Firstly, Gabor filter and Adaboost are introduced which are the main tools in the method. Then Based on the concept of IntraPersonal and ExtraPersonal space, AdaBoost is taken to select the effective Gabor features and to build the strong classifier. In order to find the law of the distribution of Gabor features which discriminate different people most effectively, some analysis is done on the sequence of Gabor features. As the result, the direction of how to select features can be found.Finally, the strong classifier built by AdaBoost is taken to recognize different faces. Based on the AdaBoosted Gabor features, LDA is taken on to train model and then to classify different faces. In the following, some experiment was done to test the generalization ability of the method using LDA and Gabor features selected by AdaBoost, and this method is compared with the method using LDA and Gabor features without selection. From the experiment result, we can draw this conclusion that the generalization ability is in inverse proportion to the number of Gabor features. |
