Jetson 기반 엣지-클라우드 오프로딩을 이용한 얼굴감정인식 프레임워크

초록

This paper proposes an adaptive Edge-Cloud framework that performs real-time facial emotion recognition (FER) on Jetson-class edge devices and performs selective offloading with a highprecision cloud model only in difficulty samples. At the edge, low-latency primary inference is performed using MobileNetV3-Small (FP32), and offloading is determined based on prediction reliability, entropy, and latency. In the cloud, ResNet-50(FP32) is used to supplement the accuracy of the high-level samples. It discloses the experimental protocol and implementation details (pre-processing, learning, and inference pipelines) on the DFEW and RAVDESS dataset, and shows the possibility of improving the accuracy-delay Pareto by simultaneously achieving improved accuracy compared to Edge-only and latency and bandwidth reduction compared to Cloud-only.