Facial Landmark-Based Face Shape Classification: A Lightweight Approach for Real-Time Applications

A lightweight, landmark-driven framework for real-time face shape classification is proposed, eliminating the need for deep neural architectures. Fourteen critical facial points are extracted using MediaPipe Face Mesh, from which five scale-invariant geometric features—jawline width, cheekbone width, face length, forehead width, and mouth width—are computed to form a concise feature vector. A shallow decision tree (maximum depth $=6$) yields an overall accuracy of 91.7 % across six canonical face shape categories on a 14 378-image dataset, with per-frame processing time under 2 ms and sustained throughput of $25-28 \text{FPS}$. The method's low computational and memory requirements render it suitable for deployment in resource-constrained environments such as mobile augmented reality, virtual try-on services, and telehealth applications.