Real-Time Facial Expression Recognition using Optimized Convolutional Neural Networks

Facial expression recognition (FER) is a key element for human-computer interaction, affective computing, and behavioral analysis. This paper proposes an optimized convolutional neural networks (CNN) architecture to perform real-time FER, attempting to find a good trade-off between high classification accuracy with low inference latency. The model is trained on benchmark datasets including FER-2013 and outdoors samples with significant preprocessing, and data augmentation to help the model generalize variation in lighting, pose, and occlusion inclusion. Results of the experiments show the proposed architecture achieves the best accuracy to inference latency ratio while requiring significantly less compute time, characterizing it for interactive systems, and edge-based applications. Future work will explore working with temporal modeling and attention mechanisms to obtain improvements.