Lightweight Convolutional Neural Network with Efficient Channel Attention Mechanism for Real-Time Facial Emotion Recognition in Embedded Systems

This paper presents a novel deep neural network for real-time emotion recognition based on facial expression measurement, optimized for low computational complexity, called Lightweight Expression Recognition Network (LiExNet). The LiExNet architecture comprises only 42,000 parameters and integrates convolutional layers, depthwise convolutional layers, an efficient channel attention mechanism, and fully connected layers. The network was trained and evaluated on three widely used datasets (CK+, KDEF, and FER2013) and a custom dataset, EMOTION-ITCH. This dataset comprises facial expressions from both industrial workers and non-workers, enabling the study of emotional responses to occupational stress. Experimental results demonstrate that LiExNet achieves high recognition performance with minimal computational resources, reaching 99.5% accuracy on CK+, 88.2% on KDEF, 79.2% on FER2013, and 96% on EMOTION-ITCH. In addition, LiExNet supports real-time inference on embedded systems, requiring only 0.03 MB of memory and 1.38 GFLOPs of computational power. Comparative evaluations show that among real-time methods, LiExNet achieves the best results, ranking first on the CK+ and KDEF datasets, and second on FER2013, demonstrating consistent performance across these datasets. These results position LiExNet as a practical and robust alternative for real-time emotion monitoring and emotional dissonance assessment in occupational settings, including hardware-constrained and embedded environments.