OPTIMIZED AI MODELS FOR FACIAL LANDMARK ESTIMATION ON SMARTPHONES AND IOT DEVICES

Abstract

Facial landmark estimation plays a pivotal role in diverse real-time applications, including identity authentication, expression recognition, and augmented reality. With the proliferation of resource-constrained devices like smartphones and IoT nodes, optimizing AI models for efficiency, speed, and accuracy is critical. This research examines the development and deployment of lightweight yet accurate deep learning models tailored for mobile and edge computing environments. Leveraging knowledge distillation, pruning, quantization, and architecture design such as MobileNetV2 and BlazeFace, this paper evaluates state-of-the-art strategies to balance precision and computational feasibility in facial landmark estimation on low-resource platforms.