无线电工程2025，Vol.55Issue(8)：1547-1559,13.DOI:10.3969/j.issn.1003-3106.2025.08.001
基于3D可解释性神经渲染的单样本人脸重演方法
3D Explainable Neural Rendering Based Single-sample Face Reenactment
摘要
Abstract
In the field of controllable face generation,face reenactment technology stands out as a pivotal research.Its objective is to utilize a given driving face image or video frame to drive the source face image,thereby enabling the precise and controllable generation of facial expressions and postures.This technology demands that the generated outcomes preserve the identity characteristics of the source face image while closely mimicking the expression postures of the driving face image.For the single sample face reenactment task,the sole reliance on 2D face images from a single viewpoint often leads to an inadequate description of facial information.Conventional approaches encounter challenges in accurately maintaining the consistency of face identity and expression postures when generating face images with significant posture variations.To tackle this issue,a novel 3D Explainable Neural Rendering Based Single-sample Face Reenactment(3D-ENS)approach is presented.This method explicitly models the fixed 3D face structure and texture information within the neural network for the entire face reenactment video generation stage.This aims to guarantee the consistency of face identity and the stability of expression and pose variations in reenactment results.Secondly,a neural texture completion network is constructed based on this.High-quality facial texture reconstruction is accomplished via a multiscale feature learning mechanism.Thirdly,a background motion estimation network is proposed to predict the background of the driven face image.Then,the background information is integrated with the completed facial Neural Texture Rendering(NTR)result.Finally,a facial landmark detection model is employed to impose facial consistency constraints,further enhancing the apparent consistency of the model.Experimental results on mainstream benchmark datasets and data in the wild verify that the proposed method exhibits excellent identity preservation capabilities.It can effectively cope with complex scenarios featuring facial posture changes,thus offering a new solution for the practical application of face reenactment.关键词
人脸重演/3D人脸建模/神经渲染/可控人脸生成Key words
face reenactment/3D face modeling/neural rendering/controllable face generation分类
信息技术与安全科学引用本文复制引用
李碧莹,刘智威,曾豪,王金桥..基于3D可解释性神经渲染的单样本人脸重演方法[J].无线电工程,2025,55(8):1547-1559,13.基金项目
国家自然科学基金面上项目(62276260) General Program of National Natural Science Foundation of China(62276260) （62276260）
