Incremental Sequential Three-Way Decision Using a Deep Stacked Autoencoder.

Most traditional face recognition classifiers attempt to minimize recognition error rate rather than misclassification costs, which is unreasonable in many real world applications. On the other hand, many facial images are usually unlabeled, and the label process may result in high costs. Considering imbalanced misclassification costs and the hardship of gathering sufficient labeled images, an incremental sequential three-way decisions (3WD) model for cost-sensitive face recognition is proposed, in which a deep stacked autoencoder (DSAE) is used to extract an efficient deep feature set. The model takes full account of the costs of obtaining labeled data in real life. In addition, the model incorporates the boundary decision into the process of making decision, leading to a delayed decision with insufficient labeled images, which simulates the decision-making process from a small amount to a large amount of data. In summary, the model aims to select an optimal decision step so as to gain the desirable recognition results with the least amount of data. This strategy is applied to two facial image databases, which validate the effectiveness of the proposed methods.