A Novel Face Emotion Recognition Based on Lite ResNet-50 for Embedded Systems

Abstract

 Face Emotion Recognition (FER) is essential for improving human-computer interaction. Although deep learning has greatly enhanced the accuracy of FER, deploying these models on embedded devices is challenging due to limitations in computational power and memory. This study proposes a lite and effective FER system for real-time implementation in resource-constrained environments. Lite ResNet-50 and MobileNetV2 were compared for the classification of neutral and angry emotions using a carefully selected dataset. The focus was on measuring accuracy, inference speed, and resource efficiency. Real-time testing is also con-ducted to confirm their applicability in practice. The results show that Lite Res-Net-50 outperforms MobileNetV2 in all key areas, achieving an accuracy of 99.8, inference speeds of 229 ms and with 2 FPS. These findings establish Lite Res-Net-50 as the optimal choice for FER on embedded devices, bridging the gap be-tween deep learning advancements and real-world deployment to improve hu-man-computer interaction.