海量数据小训练样本集环境下的2D+3D人脸表情融合识别研究

In view of efficient, effective and natural intelligent human computer interaction with low cost, the research on facial expression recognition under complicated situations, e.g., variant illumination and pose, is developed. It is expected to achieve the following goals: based on the simulation and fusion of the multiple information obtained from the stereo vision structure and 3D perception mechanism of insect compound eyes, the anatomical structure and motion characteristics of face, and the human learning mechanism by one-shot or just a handful of examples, a single web camera driven facial expression recognition system is proposed based on the establishment of a 2D+3D facial expression fusing recognition scheme in the analysis/synthesis framework. This system can run in real-time as well as having high recognition score. To alleviate the difficulties of the statistical learning with sparse samples, high accuracy and high efficiency, the following problems will be studied thoroughly, i.e., the 3D face reconstruction based on stereo vision, the facial expressive animation with 3D facial model, the extraction and fusion of static and dynamic facial expression information, and the combination of deep learning and Bayesian Program Learning. Moreover, key technologies will be established and a practice real-time and robust facial expression recognition system will be constructed to provide a concrete foundation for applications.

本项目以便捷有效、低成本代价且具备友好交互能力的智能人机交互为研究对象，系统开展可变光照和可变姿态等复杂条件下的人脸表情识别研究。冀此研究达到如下目的：充分模拟和融合昆虫复眼的立体视觉结构和三维感知机理、人脸的解剖结构和运动特性以及人类的一次性和少样本学习机制等多种生理信息，设计并实现基于分析/合成框架的2D+3D人脸表情融合识别方案，实际构建出基于单个网络摄像机的、具有高识别率的实时人脸表情识别系统。针对系统实现过程中所面临的海量数据小训练样本集下的统计学习以及高精度和高效率难题，拟从系统的角度，对基于立体视觉的人脸三维重建、基于人脸三维模型的人脸表情动画合成、人脸表情静态/动态信息的提取和融合、深度学习和贝叶斯程序学习的结合等诸问题进行深入研究，确立与之相关的关键技术并实际构建出实时鲁棒的人脸表情识别系统，为研究走向实用化奠定基础。

本项目以便捷有效、低成本代价且具备友好交互能力的智能人机交互为研究对象，系统开展可变光照和可变姿态等复杂条件下的人脸表情识别研究。此研究达到了如下目的：充分模拟和融合昆虫复眼的立体视觉结构和三维感知机理、人脸的解剖结构和运动特性以及人类的一次性和少样本学习机制等多种生理信息，设计并实现了基于分析/合成框架的2D+3D人脸表情融合识别方案，实际构建出了基于单个网络摄像机的、具有高识别率的实时人脸表情识别系统。针对系统实现过程中所面临的海量数据小训练样本集下的统计学习以及高精度和高效率难题，从系统的角度，对基于立体视觉的人脸三维重建、基于人脸三维模型的人脸表情动画合成、人脸表情静态/动态信息的提取和融合、深度学习和贝叶斯程序学习的结合等诸问题进行深入研究，确立了与之相关的关键技术并实际构建出了鲁棒的人脸表情识别系统。