Research on Lightweight Facial Landmark Prediction Network

Abstract

Facial landmarks, as direct and reliable biometric features, are widely utilized in various fields, including information security, public safety, virtual reality, and augmented reality. Facial landmarks, which are discrete key points on the face, preserve expression features and maintain the topological structure between facial organs. Fast and accurate facial landmark prediction is essential in solving computer vision problems involving facial analysis, particularly in occlusion scenarios. This research proposes a lightweight facial landmark prediction network for occluded faces using an improved depthwise separable convolutional neural network architecture. The model is trained using 30,000 images from the CelebA-HQ dataset. The model is then tested under different occlusion ratios, including 10–20%, 30–40%, 40–50%, and 50–60% random occlusion, as well as 25% center occlusion. Using 68 facial landmarks for occlusion prediction, the proposed method always achieved significant improvements. Experimental results show that the proposed lightweight facial landmark prediction method is 1.97 times faster than FAN* and 1.67 times faster than ESR*, while still achieving better prediction results with lower NMSE values across all tested occlusion ratios for both frontal and profile faces.