Enhancing Facial Reconstruction Using Graph Attention Networks

Traditionally, research on three-dimensional (3D) facial reconstruction has focused heavily on methods that use 3D Morphable Models (3DMMs) based on principal component analysis (PCA). Because such methods are linear, they are robust to external noise. The PCA method has limitations when restoring faces that deviate from the training data distribution, particularly when recovering fine details. By contrast, restoration methods utilizing Graph Convolution Networks (GCN) offer the advantages of non-linearity and direct regression of vertex coordinates and colors. However, GCN-based approaches can be prone to overfitting, making them less stable. This study presents a face restoration approach that aims to regress the vertex coordinates and colors of a 3D face model directly from a single wilds 2D facial image. This method demonstrates greater stability and higher accuracy compared to conventional techniques. In addition, Graph Attention Networks (GAT) enhance the restoration performance while separating the networks responsible for facial shape and color, reducing noise caused by interference between different data attributes. Through experiments, we demonstrate the most optimized network structures and training methods and demonstrate improved performance compared to existing approaches.