Pre-trained Deep Learning Models for Facial Emotions Recognition

The present study is devoted to a comparative analysis of known pre-trained models of deep learning neural networks used to facial emotions recognition. The aim of the study is to select appropriate lightweight models to be used for offline facial emotions recognition of the students during their semestrial learning. The basic facial emotions can be grouped as positive (happiness and surprise), neutral, and negative (fear, anger, sadness and disgust). Based on these three groups of emotions corrective actions by adapting and personalizing the lecture material are undertaken.

I. Introduction

Facial emotions are one of the most informative means of assessing the attitude of a person to the environment, a topic, object, commercial product and more. Currently, technologies such as facial recognition, fingerprint recognition, voice command recognition or the integration of voice virtual assistants such as Siri, Alexa and Google Assistant are part of the software of almost all smartphones or devices available on the market. Part of these technologies based on deep neural networks are intended for image and human face recognition and a specific part of them to facial emotions recognition (FER). In recent years, there has been a rapid development of convolutional neural networks (CNN) for facial recognition and facial emotion recognition. Thanks to the efforts of the scientific community, many trained models of neural networks for FER such as DeepFace, VGG16, VGG19, ResNet, etc., are available for non-commercial use. In the present study, an analysis of several existing FER models was made in order to select an appropriate model to be used to analyze students' emotions during their studies. As on the basis of their emotions (positive, neutral or negative) actions to personalize and individualize the lecture material to the particular student are performed. The structure of this report is as follows: in section 2 the emotions and their application in different subject areas are discussed; section 3 describes the process of recognizing facial emotions through CNN; the most common datasets for FER are discussed in section 4; in section 5 includes an analysis of trained models for FER; section 6 compares the results of predicting the emotions of students from 2 selected models; conclusions have been drawn in section 7.