Realizing High Fidelity 3D Face of Images Using Self Encoding Networks
Hao Xiao, Geguo Du, Yingyu Zhang, Wenjun Yang · 2023
Single-view 3D face reconstruction, a process of transforming 2D facial images into 3D models, is pivotal in diverse fields such as face recognition, gaming, and entertainment. This paper presents an advanced method for 3D face reconstruction using a self-coding network. Our approach focuses on extracting detailed 3D facial features from 2D images with a feature encoder, followed by the generation of a comprehensive 3D face model. Concurrently, we employ an expression prediction network to analyze and integrate facial expression geometry, refining the model's expression feature parameters. Additionally, an albedo modification module is introduced to distinguish between illumination and facial geometry, enhancing the accuracy of texture details. To validate the method's effectiveness, we have devised multi-level loss functions for detail reconstruction. These functions serve as self-monitoring constraints and are rigorously evaluated against established methods using a dataset, confirming the proposed method's reliability and superiority in achieving high-fidelity 3D face reconstructions.