Kernel Machine for Fast and Incremental Learning of Face
Woo-sung Kang, Jin Young Choi · 2006 SICE-ICASE International Joint Conference · 2006
This paper proposes a novel method for improving training speed and incremental learning on multi-class classification such as face recognition. In existing system, the training time of multi-class SVM using binary classifier increase rapidly due to the repeated use of data with the increase of training data and the number of class. In the case of large data set, this leads to difficulty of training due to limited resource in practical application. Thus, we propose a new multi-class classification method based on support vector domain description (SVDD) that can learn incrementally by using just one class data for training a added person. The proposed method can reduce training time and computational load by avoiding the repeated use of data. To verify the performance of the proposed method, experiments are carried out in comparison with three other methods: one-against-all, one-against-all and neural network. The experimental results show that the proposed method is more adequate than other method for multi-class problem with respect to training speed and computational load