High-Resolution Multi-Level Feature Fusion Network for Robust Facial Landmark Detection

Abstract

Facial keypoint detection is relatively easy, but not completely accurate. In practical applications, occlusion, pose variations, and image noise are obstacles to accuracy. To achieve more accurate and stable keypoint detection, this paper will address the aforementioned issues by constructing a high-resolution neural network. The network uses a multi-layer feature fusion structure to extract global facial features and local details at different scales. In order to balance the trade-off between overall smoothness and high-resolution accuracy in model optimization, a new combined loss function was adopted. It will be more robust to blurred and occluded samples. 300-W, COFW, and WFLW are public data experimental datasets. The normalized mean error (NME) on 300-W is 2.89, on COFW is 3.56, and on WFLW is 3.19, all three methods have set new records. Common interferences such as blur and noise have been better resisted. Ablation experiments demonstrated the fusion strategy and loss components. Qualitative visualizations and application cases indicate that the system can perform well in various environments and under different conditions. These two methods are both practical and reliable, and can be used in many fields, such as mobile health monitoring, biometric authentication, and human-computer interaction.