Integrating Multimodal Data for Deep Learning-Based Facial Emotion Recognition

Abstract

With the rapid development of neural networks, emotion recognition has become a research area of great concern. It has important applications not only in marketing and human-computer interaction but also holds significant importance for improving emotional computing and user experience. This paper studies various methods for emotion recognition in images and videos, utilizing convolutional neural networks (CNN), multi-layer perceptron (MLP), and fusion models. The Facial Expression Recognition 2013 (FER2013) image dataset and the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) audio and video dataset serve as the basis for this study. The experimental results indicate that ResNet18 outperforms others in image emotion recognition, attributed to its residual block design and the incorporation of regularization techniques. In the realm of video emotion recognition, the audio model based on MLP demonstrates a superior ability to identify emotional information. Although the fusion of image and audio models theoretically could enhance accuracy, the randomness of video frames prevents the fusion model from achieving the desired effect. Future research might further explore the application of time series models in video emotion recognition to capture continuous emotional changes within videos.