面向三维人脸的美学特征分析与智能感知研究

Unlike 2D images that suffer from the viewpoint limitations thus fail to represent the true structure of faces, the research based on 3D faces can investigate the nature and components of facial aesthetics more completely and accurately. The challenges lie in the lack of discriminative aesthetic representation caused by the complexity of 3D feature extraction, as well as the limited aesthetic computational performance caused by the subjectivity of visual perception and the scarcity of training data. To address these problems, this project aims to construct an accurate feature representation through handcrafted rules and data mining, and then develop a label distribution-based aesthetic prediction model by the rating label distribution. In this project, 3D faces will be used as the experimental data. In order to normatively describe the facial aesthetics, the aesthetic-aware face representation based on its structural characteristics will be explored in both rule-driven and data-driven manner. Inspired by different quantitative forms of manual labels, the aesthetic computational methods which are applicable to small datasets will be studied to achieve the automatic prediction of aesthetic perception. The above theoretical results can be used in medical cosmetology, and finally facilitate the development of a computer-aided system for craniomaxillofacial surgery planning and simulation. The project is expected to bring new ideas to facial aesthetics studies from multiple disciplines, but also provide a reference for other research on image analysis and understanding, and biometrics.

面向三维人脸的美学研究摆脱了二维视角难以表达人脸真实结构的限制，能更加完整精确地探索面部美学的本质与构成。其难点在于三维特征提取的复杂性使得人脸表征的美学鉴别力不足，同时视觉感知的主观性和训练数据的稀缺性导致美学计算性能受限。针对这些问题，本项目从准则设计和数据挖掘两个角度构建清晰准确的特征描述模型，利用人工审美标注的分布特点建立标记分布范式下的感知预测模型。项目以三维人脸数据作为研究对象，依据准则驱动和数据驱动两种模式，挖掘适应于面部三维结构的美学特征表达，实现对人脸美的规范描述；结合人工标注的多种量化形态，研究适应于小样本数据集的美学计算建模方法，实现对审美倾向的智能预测；将理论成果应用于医学整容实践，实现一个计算机辅助的颅颌面手术设计与仿真系统。本项目有望为跨学科领域的人脸美学研究带来新思路，还能为其他图像分析理解和生物特征识别研究带来借鉴。

面向三维人脸的美学研究摆脱了二维视角难以表达人脸真实结构的限制，能更加完整精确地探索面部美学的本质与构成。其难点在于三维特征提取的复杂性使得人脸表征的美学鉴别力不足，同时视觉感知的主观性和训练数据的稀缺性导致美学计算性能受限。针对这些问题，本项目开展了三维人脸特征分析、人脸美学智能感知、整容手术仿真应用等相关研究，获得了如下进展。提出了基于深度语义分割的三维模型标志点检测方法；研究了基于点云和网格数据的特征学习方法，并作用于三维人脸美学表征；在准则驱动和数据驱动两种模式下，挖掘适应于面部结构的美学特征表达；构造标记分布范式下的美学量化形态，搭建轻量化美学预测模型；针对三维模型碎块，提出了基于平衡聚类树的断裂面配准和基于多视角的断裂线提取；面向人脸表情应用，实现了表情动画生成和端到端表情识别；将上述理论成果应用于医学整容实践，开发了一款计算机辅助的颅颌面手术设计与仿真系统。本项目为跨学科领域的人脸美学研究带来了新思路，也为其他图像分析理解和生物特征识别研究提供借鉴。