A Deep Convolutional Neural Network With Fuzzy Rough Sets for FER

Xiangjian Chen,Di Li,Pingxin Wang,Xibei Yang

doi:10.1109/access.2019.2960769

Abstract

Existing facial emotion recognition methods do not have high accuracy and are not sufficient practical in real-time applications. We introduce type 2 fuzzy rough sets to develop a Type 2 Fuzzy Rough Convolutional Neural Network, as type 2 fuzzy rough sets form a suitable mathematical tool to characterize uncertainty of classifification. Based on the type 2 fuzzy rough sets theory, we construct an optimization objective for training CNNs by minimizing fuzzy classification uncertainty, and present the defifinition and optimization of type 2 fuzzy rough loss, which can be achieved by better performance. This method could reduce the uncertainty in terms of vagueness and indiscernibility by using type 2 fuzzy rough sets theory and specififically removing noise samples by using CNN from raw data. And finally, compared the proposed method with other feature extraction and learning techniques based on Algorithm Adaption k-Nearest-Neighbors. Experimental results demonstrate that type 2 fuzzy rough sets convolutional neural network could achieve better performances comparing with other methods.

Highlights

The expression is a direct way of transmission of human inner emotions and plays a subtle role in interpersonal communication
FRAMEWORK OF THE PROPOSED METHOD The deep convolutional neural network has been broken through the graph classification questions, and the work has solved the problem of the classification of 1,000 categories of pictures in the ImageNet data set
In the fuzzy decision-making stage, the pre-trained type 2 fuzzy rough sets convolutional neural network is used as a depth feature extractor, and fuzzy expression classification based on special characteristics is constructed, which is used for fuzzy decision-making

Summary

INTRODUCTION

The expression is a direct way of transmission of human inner emotions and plays a subtle role in interpersonal communication. As a mathematical tool to effectively deal with the inconsistency of label information, has been widely discussed and applied in fuzzy classification problems It is often used for feature selection and feature weighting of fuzzy classification problems, and provides a good idea for the training. The fuzzy rough set model has been successfully applied to heterogeneous attribute reduction [16], active learning [17], [18], build a robust classification method [19], fuzzy rule extraction [20], MR image segmentation [21], integrated learning [22], and Shan et al [23] proposed covering-based general multigranulation intuitionistic fuzzy rough sets and corresponding applications to multi-attribute group decisionmaking In other fields, it shows excellent performance. Experiment results and analysis are presented in Section ‘‘Experiments and Validation.’’ Lastly, Section ‘‘Conclusions’’ presents our conclusions

FRAMEWORK OF THE PROPOSED METHOD

TYPE 2 FUZZY LOSS FUNCTION

CONCLUSION