非约束环境下人脸多属性分析的理论与方法研究

Facial attribute recognition is to estimate the facial attributes, such as gender, expression, ethnicity and age etc., based on the extracted facial features. It is widely used in the fields of face image search, suspicious person monitoring, etc. Existing facial attribute recognition methods usually extract low-level visual features and do not use the correlations between the attributes sufficiently. Moreover, these methods are not robust to environment change due to the effects of illumination, pose, occlusion, image quality, etc. Focusing on the scientific problem of facial attribute representation and recognition in unconstrained environment, this project aims to investigate the theory and methods of face image normalization in unconstrained environment, facial attribute feature learning and the usage of correlations between facial attributes, and to break through the key technologies of face image normalization based on adaptive multi-column stacked sparse denosing autoencoder, multi-attribute feature learning based on deep convolutional neural networks, and classifier design by incorporating the correlations between the attributes and fusing multi-attribute classification results. This project will develop a real-time face search system based on multi-attribute recognition, which can reach the state-of-the-art performance and lay the foundation for the significant security applications like suspicious person monitoring under large-scale video surveillance.

人脸属性识别是指基于面部特征来对人脸的性别、表情、种族、年龄等属性做出判断，在人脸图像检索、可疑人员监控等多个领域具有广阔的应用前景。现有人脸属性识别方法通常只提取低层图像特征、对属性之间相关性的利用不够充分；另外由于光照、角度、遮挡、图像质量等因素的影响，现有方法对环境变化不够鲁棒。本项目将围绕非约束环境下人脸属性的表示与识别这一关键科学问题，针对非约束环境下的人脸图像正规化、人脸属性特征的学习、人脸属性之间相关性的利用等理论与方法开展研究，重点突破基于自适应多列栈式稀疏去噪自编码器的人脸图像正规化、基于深度卷积神经网络的多种人脸属性特征学习、融合属性之间相关性的分类器设计及多属性分类结果融合等关键技术。项目将开发基于多属性识别的人脸图像实时检索系统，其性能达到国际同类算法的领先水平，为大规模视频监控网络下可疑人员监控等重大安防应用奠定基础。

对人脸属性的分析，首先是将人脸区域在图像或视频帧中定位，然后再进行人脸关键点检测、人脸图像正规化，最后完成人脸属性特征的分析和判别。不同光照、角度、遮挡和图像质量等因素对这些任务带来了挑战。本项目针对人脸属性识别的各个模块以及人脸属性之间相关性的利用等多个方面进行了系统地研究，具体包括：（1）基于多尺度信息的人脸检测方法；（2）人脸关键点检测方法；（3）不同姿态的人脸图像正规化方法；（4）基于数据增强和局部信息的特征提取模型训练方法；（5）基于跨域和含有弱标签数据集的深度卷积网络构建、特征分析方法。取得的重要成果包括：（1）提出基于分层全卷积网络的人脸检测方法；（2）提出基于双层深度初始化网络人脸关键点定位方法；（3）提出基于自适应摄像头标定的人脸图像正规化方法；（4）考察了基于图像处理、关键点扰动、3D可变模型（3DMM）扩展等不同的数据增强方法，增强人脸特征的描述能力。（5）提出了隐式人脸模型和自引导的联合贝叶斯方法，提高人脸属性和身份等特征在不同数据集的可区分性。上述研究成果推进了人脸多属性识别的发展，并促进了人脸属性识别与人脸识别系统的相关推广与应用。