DS-CTAN for Lightweight Facial Expression Recognition in Edge Systems
Abstract
This paper presents DS-CTAN, a lightweight two-stream network for facial expression recognition in edge-based distributed systems. Designed for heterogeneous devices such as mobile phones, Jetson, and Raspberry Pi with TPU, it combines spatial feature extraction with temporal attention to capture dynamic facial movements. Using transfer learning and knowledge distillation, the model achieves high accuracy with low computational cost. Evaluated on multiple public datasets and a self-built cross-domain benchmark, DS-CTAN enables real-time, on-device inference with fast response and low power consumption. By performing inference locally, it reduces latency and data transmission, making it suitable for distributed AI applications like smart environments, human-computer interaction, and edge-driven affective computing. The framework provides a practical solution for deploying efficient vision models in resource-constrained, distributed architectures.