A Design and Implementation of Evaluation System for Lightweight Facial Expression Recognition Models in Embedded Environments

In this paper, we propose and implement an evaluation system designed for the efficient verification of the practical applicability of lightweight facial expression recognition (FER) models in embedded environments. While various lightweight FER models have been introduced, their actual performance evaluation and verification in resource-constrained embedded environments, though crucial, remain an underexplored area of research. Therefore, this study focuses on addressing this gap in existing research regarding the applicability of lightweight models in embedded environments. The proposed evaluation system is designed to quickly assess the performance and applicability of lightweight FER models in embedded environments and features a modular structure that allows for the systematic measurement of model accuracy, processing speed, and resource usage. Using the developed evaluation system, we assessed the performance of key lightweight FER models, including Poster V2, DAN, and Multi-task EfficientNet B2, on an NVIDIA Jetson Orin Nano board using the AffectNet 8 Emotions dataset. Experimental results showed that while the evaluated models exhibited similar average accuracy ranging from approximately 61-63% on the AffectNet 8 Emotion dataset, significant differences were observed among models in terms of processing speed and memory usage. Specifically, the DAN model achieved the fastest inference speed compared to other models but simultaneously recorded the highest memory usage, suggesting that a significant trade-off between performance and resource efficiency must be considered when selecting models for embedded environments. This study makes important contributions by proposing a practical framework for evaluating FER models in embedded system environments and providing benchmark data for lightweight models on actual hardware. These results are expected to provide practical insights for selecting lightweight FER models optimized for resource-constrained real-world applications, such as robotic platforms, and for future research aimed at improving their efficiency.