Construction Method of 3D Face Point Cloud Dataset Based on Data Enhancement Mechanism
Lixia Liu · 2024
3D facial point cloud data can provide more abundant facial structural information, but the acquisition and processing of such data are relatively complex, requiring specific hardware devices and software support. The scale of existing 3D facial datasets is typically small, which limits the training effect of deep learning-based recognition networks and ultimately leads to a decrease in recognition accuracy. This paper proposes a method for constructing a 3D facial point cloud dataset based on a data enhancement mechanism, aiming to address the challenges of data collection and model training in 3D facial recognition technology. Firstly, a relatively small amount of facial data is collected, and facial localization and key point recognition are performed using a face detector and key point detector, and the depth maps is converted into 3D point cloud data. Then, various data enhancement mechanisms are introduced to expand the scale of the dataset, including data enhancement based on bending energy, rotation transformation, and Gaussian noise, which not only increase the diversity of the data but also improve the generalization ability of the model under different lighting and poses, thus constructing a 3D facial point cloud dataset with millions of IDs. Experimental results show that the 3D facial point cloud dataset constructed through the proposed method achieves higher recognition accuracy on various network models such as SpiderCNN, KPConv, PCN, PRNet, and PointNet++ compared to popular datasets like Bosphorus 3D, CASIA 3DFE, FaceScape, and Look3Dface, verifying the effectiveness of utilizing a small amount of face data and expanding a large dataset through data enhancement mechanisms. Additionally, the ablation experiments also confirm the effectiveness of data enhancement mechanisms based on bending energy, rotation transformation, and Gaussian noise, indicating that these mechanisms can significantly improve the quality of the dataset and model recognition accuracy.