High-Fidelity 3D Face Reconstruction from a Single Image via Gaussian Representation and Depth-Guided Refinement

3D facial reconstruction has been extensively applied across various domains; however, existing methods continue to face challenges in balancing computational efficiency with geometric accuracy. Traditional techniques depend on expensive hardware and are often time-consuming, while learning-based approaches, despite notable advancements, still struggle with preserving fine details and identity-specific features. To address these limitations, a single-view 3D face reconstruction algorithm is proposed. The method employs a UNet-based architecture to project a monocular facial image into a 3D Gaussian representation, enabling high-quality novel view synthesis through Gaussian Splatting. To mitigate the scarcity of multi-view facial datasets, a large-scale dataset is synthesized using the pre-trained 3D GAN model EG3D, providing high-quality supervision for training. A refinement strategy is further introduced, incorporating face super-resolution and a six-channel image loss function to enhance visual clarity and detail fidelity. Additionally, a depth-guided refinement scheme is proposed, leveraging precise depth maps acquired from structured light scanning systems to optimize the Gaussian field for accurate geometry and fine texture reconstruction. Experimental results show that the proposed method achieves competitive performance in both reconstruction quality and computational efficiency.