Gaussian Mixture Model Based Cost Sensitive Sequential Three-Way Decision For Face Recognition

In the era of artificial intelligence,face recognition technology has reached a high maturity.However,most traditional face recognition systems take reducing the number of misclassifications and improving the classification accuracy as the main goal,but ignore the fact that different misclassifications may cause different costs in reality.Moreover,training an accurate face recognition classifier requires a large number of labeled samples.In practical applications,the number of labeled samples is often very small,and making decisions directly when the labeled samples are insufficient is likely to cause high classification misclassification costs.In order to solve this problem,considering that useful information is constantly increasing in reality recognition,this paper proposes a sequential three-way decision for face recognition method based on Gaussian mixture model,which is used to solve the cost-sensitive face recognition problem when there are insufficient labeled samples.Firstly,using the idea of semi-supervised learning,Gaussian mixture model is used to model all the face image data,and face image classification is realized based on EM algorithm.Secondly,combined with the idea of self-training algorithm,GMMTraining algorithm is proposed to construct a dynamic incremental labeled sample set,which makes full use of unlabeled samples to help the trained classifier to better represent test samples and obtain a more ideal recognition effect.Then,in the process of face image recognition and classification,cost sensitive and three-way decision methods are introduced,which not only consider to improve the classification accuracy,but also achieve the balance between the classification accuracy and the cost of classification error.Finally,this paper constructs an incremental sequential three-way decision model from coarse information granularity to fine information granularity,and measures the performance of face recognition system with the total cost.Comparison experiments on the classification accuracy,decision cost and total cost of the proposed method are carried out on EYale B,PIE and AR face datasets,and the effectiveness of the proposed method is verified by experiments.