Convolutional Neural Network-Based Classification of Facial Emotional Expressions and Computational Complexity Analysis

Abstract

This study presents a novel convolutional neural network (CNN) model and a detailed complexity analysis with other models available in the literature for an accurate classification of facial emotional expressions. Human beings have been defined by seven basic emotions, which are anger, fear, happiness, sad, contempt, disgust, and surprise. Model accuracy plays an important role in emotion detection studies with deep neural networks, as the high model accuracy is directly related to the accuracy of the predicted emotions. A 23-layer CNN model was created that classifies 7 different emotions. The CNN model we trained with the FER2013 dataset has a higher accuracy performance than other studies trained with the same dataset in the literature. The accuracy performance of our CNN model is 98.83% in training data and 83.52% in validation data. The complexity of the algorithms used in other studies is compared with the proposed study. Although the accuracy performance of our CNN model is higher than other studies in the literature, the complexity of our model is also higher than most other studies. The CNN model we obtained is used in an algorithm that we have created to increase the efficiency of online courses, which performs sentiment analysis 4 times per second.