Semi-Supervised Incremental Three-Way Decision Using Convolutional Neural Network

This paper aims to develop a novel cost-sensitive face recognition framework, which can gain the desirable recognition results with the least total cost. By combining two recently rising techniques: deep convolutional neural networks (CNNs) and sequential three-way decision (3WD) method, our framework can automatically label new samples and incorporates the delayed decision into decision-making process. We first explore the semi-supervised face recognition method in the case of the scarcity of labeled training data. By learning the class estimation and the deep convolution feature extraction of the unlabeled data jointly, the CNN trained by both labeled training data and unlabeled data is generated. Then, rather than getting a lower recognition error rate, we focus on seeking the minimum cost of misclassification at each decision step. For this purpose, we introduce the method of sequential 3WD in our cost-sensitive face recognition framework, which take each iteration of semi-supervised learning as a decision-making step. When there are insufficient labeled samples, a delayed decision will be adopted to reduce the decision cost. Finally, the test cost is also considered in the decision-making process, and the sum of misclassification cost and test cost is taken as the total cost. Using the total cost as the objective function, optimizing the performance indicators, training to get the classifier with the smallest total cost. In short, the model strives to get an optimal decision step, so that the reliable identification result can be obtained with only a small number of labeled data. The work value of this paper is to prove the effectiveness of our method in two face datasets.