Three-Dimensional Face Reconstruction Based on Single Image Features
Yankun Feng, Kai-yu Wu · 2025
To address the challenges of low landmark detection accuracy and insufficient high-frequency detail restoration in 3D caricature face reconstruction from a single image, this paper proposes a two-stage method combining multi-scale feature fusion and high-frequency information mapping. In the first stage, a landmark detector based on multi-scale channel fusion is designed to improve detection accuracy. Multi-scale features are extracted using HRNet, while the attention layer, composed of channel attention mechanisms and Swin Transformer, is employed for multi-scale channel feature fusion. To enhance the accuracy of landmark generation, the loss function consists of landmark loss and heatmap loss. In the second stage, a deformation network with a Fourier feature-sharing layer is utilized to enrich the high-frequency details of the reconstructed 3D caricature faces. The Fourier feature mapping extracts high-dimensional features, enabling the network to learn more high-frequency shape information, while the shared super-network layer accelerates network convergence and reconstruction efficiency. This method was tested on the CaricatureFace and 3DCaricShop datasets. Experimental results demonstrate that the landmark detector reduces the average detection error by 4.4%, while the deformation network decreases the mean squared error in shape reconstruction by 26% and reduces the average reconstruction time by 18%. The reconstructed 3D caricature faces exhibit exaggerated shapes and natural details.