Abstract: Reconstructing three-dimensional（3D） faces from monocular RGB images is a challenging computer-vision task. Owing to the dearth of datasets with facial-expression labels, most 3D facial reconstruction schemes lack the supervision of facial expressions,thus resulting in the inaccurate reconstruction of input facial expressions. Therefore, this study proposes a 3D facial reconstruction method based on emotional consistency. In this method, a loss of emotional perception consistency is introduced during training to selfsupervise facial emotions, thus enabling the reconstructed face to exhibit the same facial expression as the input face. Additionally, this study proposes a lightweight framework that uses MobileNetV2 to replace the deep network ResNet50 to regress face parameters and improve the inference speed of the model on the CPU side. Experimental results show that the proposed method can reconstruct a high-quality 3D face model based on a single-face image. The proposed method is superior to some mainstream face-reconstruction methods in terms of facial-expression capture and 3D face-reconstruction accuracies. Additionally,the lightweight face-reconstruction framework adopted in this study significantly improves the inference speed on the CPU side and expands the application prospects of the model in computing-resource-constrained scenarios.

Key words: three-dimensional facial reconstruction, emotional consistency, self-supervised, MobileNetV2