非约束环境下的人脸图像预处理计算模型与方法研究

Face image preprocessing is to transform a non-standard face image in unconstrained environment into the one under a unified and standard condition. It usually includes face alignment and face normalization, which can help to resolve the problems of dynamic face recognition in unconstrained environment, i.e. poor stability of facial feature and large effect of various external conditions. Due to the combined effects of many factors such as illumination, pose, occlusion, etc., the robustness to environmental change of existing face image preprocessing methods is not enough. This project aims to address this problem by investigate the theory and methods of face image preprocessing in unconstrained environment. First, to address the robustness issue of facial point detection in unconstrained environment, the face point detection algorithm based on Visibility-aware Mixtures of Part Models is investigated, which lays a good foundation for the face alignment. Then, based on the Adaptive Multi-column Deep Convolutional Neural Network, we explore face normalization method to achieve the normalization of various changes jointly, such as illumination, pose, and occlusion, etc. Finally, we establish a face recognition system on the basis of the above technologies and carry out experiments verification, to provide the key technologies for the realization of face recognition systems available in actual environments. The implementation of the project will help improve the robustness of face recognition when applying in real environments, and will effectively promote the application of face recognition.

人脸图像预处理是指将各种外界条件影响下的非标准人脸图像变换到统一的标准条件下，通常包括人脸对齐及人脸矫正，对于解决非约束环境下进行动态人脸识别所面临的人脸特征稳定性差、受各种外界条件影响大等问题具有重要意义。由于受到光照、角度、遮挡等多种因素的联合影响，现有人脸图像预处理方法对环境变化不够鲁棒。本项目将针对该问题，探索非约束环境下人脸图像预处理的理论及方法。首先针对非约束环境下人脸关键点检测的鲁棒性问题，研究基于可见感知混合部件模型的人脸关键点检测算法，为实现良好的人脸对齐奠定基础。然后，基于自适应多列深度卷积神经网络探究人脸矫正方法，实现光照、角度、遮挡等各种变化因素的联合矫正。最后在此基础上搭建人脸识别演示系统，开展实验验证，为实现实际应用环境下可用的人脸识别系统提供核心算法与技术。本项目取得的成果有利于提高人脸识别在真实应用场景中的鲁棒性，对人脸识别的实际应用起到有效的促进作用。

在实际应用环境中，由于受到光照、角度、表情、遮挡等多种因素的影响，人脸图像预处理与人脸识别面临着巨大的挑战。本项目探索了非约束环境下人脸图像预处理及人脸识别的理论及方法。主要研究内容包括：1）人脸关键点检测方法，提出基于可见感知混合部件模型的人脸关键点检测方法与基于两阶段初始化深度回归结构的人脸关键点检测方法；2）多因素联合人脸矫正方法与人脸姿态矫正方法，提出基于自适应多列深度卷积神经网络的多因素联合人脸矫正方法与基于自适应摄像头标定的人脸姿态矫正方法；3）基于隐变量分析的跨媒介人脸识别方法；4）基于关键点扰动的人脸数据生成方法；5）通过弱标注人脸图片有效训练深度卷积神经网络模型的方法；6）半监督的交叉域图像识别方法。本项目的研究成果能够进一步推动人脸识别理论的发展，为人脸识别系统的应用提供有力的技术支撑。