Abstract
After using unlabeled samples to assist training, the classifier based on semi-supervised learning sometimes not only can not improve its generalization ability, but also get a degradation of performance for the introduction of too many noise samples. To overcome this disadvantage, a new algorithm called co-training semi-supervised active learning based on noise filter is presented in this paper. This algorithm uses three fuzzy buried markov models. To avoid the rejective judgment when the classifiers do not agree with each other and the inaccurate judgment when the initial weak classifiers all agree, some human-computer interactions are actively introduced to label the unlabeled sample at suitable time. Meanwhile, the noise filter is used to filter the computer automatically labeled samples which may be noise samples. The experimental results show that this algorithm applied to facial expression recognition can effectively improve the utilization of unlabeled samples, reduce the introduction of noise samples and raise the accuracy of expression recognition.