Enhanced 3D Reconstruction of Faces Animation by Background Separation and Anisotropic Diffusion

Extensive research has propelled significant advancements in facial animation techniques, yet prevailing methods often exhibit undesirable irregularities due to animation distortion. This study introduces a novel compositional framework for face frontalization animation reconstruction, showcasing the efficacy of our proposed method. Employing a mask layer during background separation enhances the model's resilience, while a non-linear anisotropic diffusion filtering approach effectively removes noise while preserving essential edge information. Our technique prioritizes facial animation, minimizing distortions in the background. Moreover, our face frontalization method retains fundamental facial shape and texture, yielding a more realistic appearance. Extensive experiments on talking head benchmarks validate the superior efficacy of our background separation technique in controlling facial animation and mitigating background distortions, surpassing existing approaches.