Single-View 3D Face Reconstruction Method With Multi-Scale Feature Fusion Encoder
Jin Hu Sun, Yuheng Chen, Liang Dong · IEEE Access · 2025
The enhancement of detailed facial features is a critical challenge in the application of inverse engineering within the medical field. However, existing methods: Position map Regression Network (PRNet), Detailed Expression Capture and Animation (DECA) exhibit insufficient accuracy in reconstructing mid-frequency details. To address this, we propose a single-view 3D face reconstruction method based on a multi-scale feature fusion encoder, which integrates local and global features to extract mid-frequency geometric details, thereby improving the fidelity of facial expression reconstruction. Furthermore, deformable convolutions are employed to optimize the generator network, enabling precise localization of mid-frequency detail regions associated with expressive geometry. This significantly enhances the accuracy of detail reconstruction, particularly for mid-frequency geometric features.Experimental validation on the FaceScape dataset demonstrates that, compared to the DECA algorithm, our network achieves a 9.1% improvement in Chamfer Distance and a 12.3% reduction in mean normal error. The evaluation on the REALY benchmark also achieved a significant 28.0% reduction in comprehensive frontal-view error compared to DECA, confirming its capability to accurately represent mid-frequency contour details of facial structures.