关键词: 三维人脸生成, 潜变量解耦, 位置注意力, 向量离散化, 特征融合

Abstract: Abstract: Aiming at the problem of insufficient interpretability of latent features, limited disentanglement capability, and poor identity consistency in existing 3D face generation methods, a Latent Feature Enhanced for Disentanglement Variational Auto-Encoder (LFED-VAE) was proposed. Firstly, the hierarchical clustering technique was used to construct a vector discretization module to promote the potential features to absorb prior knowledge and improve the decoupling performance. Secondly, a location attention module was designed to selectively integrate the location information of potential features through element-by-element summation operation to ensure the identity consistency of generated faces. Finally, combining prior knowledge and location information, the maximum normalization technique was used to enhance the interpretability of potential features in face generation. The experimental results demonstrate that the proposed method achieves an accuracy of 95.96% in the latent feature disentanglement metric of Variability Predictability (VP). Compared with the Swap Disentangled Variational Auto-Encoder (SD-VAE), the Local Eigenprojection Disentangled Variational Auto-Encoder (LED-VAE), and the Spherical Harmonic Local Eigenprojection Disentangled Variational Auto-Encoder (SHLED-VAE), the proposed method shows improvements of 15.25, 14.62, and 12.75 percentage points, respectively. The proposed method enhances disentanglement performance while maintaining good representation and reconstruction capabilities.

Key words: 3D face generation, latent disentanglement, positional attention, vector discretization, feature fusion