Average Refined Neutrosophic Framework for Facial Emotion Recognition

Facial Emotion Recognition (FER) is a part of image processing that classifies a particular image into a primary emotion. Ekman defined six basic emotions: sadness, happiness, surprise, anger, fear, and disgust. Neutral emotions are an umbrella term for an indeterminate state or absence of precise emotion. This paper considers seven emotions, including neutral, for classification. The FER2013 dataset is benchmarked in the field of emotion recognition for face images. Most Convolution Neural Networks (CNN) models achieve close to 60% accuracy on the FER2013 dataset. This paper proposes an average refined neutrosophic framework for improving accuracy by combining neutrosophic refined sets and CNN. It represents the seven emotions using the Refined Ekman Emotion Neutrosophic Sets (REENSs) and uses machine learning algorithms for classification. The framework aids in improving the accuracy. Three different CNN-based models were used to generate the average REENS, and then it was used to check the proposed framework's performance. The proposed framework improved accuracy to 64.67%, increasing accuracy by 2%.