面向互联网图像视频的鲁棒人脸分析与识别技术研究

Along with the popularization of mobile cameras and the evolve of mobile Internet, the number of online photos and invideos increases rapidly. How to conduct efficeint and rubost face recognition in web images has become a vital problem to be solved. Web photos and videos are always taken under unconstrained and natural enviroments. Pose, illumuntation, expression, age and makeup can bring complex variations in face images. Moreover, face image in webs laways has low resolution. Faces these challenges, this project takes machine learning as main tools and conducet ressearchers from several aspects, including, reconstructing canonical view of face images, learning local descriptors, mining high level indentiy features, and developping recognition algorithmns for low resolution images. This project will develop a deep model to learn complex pose and lighting transforms in uncontrolled environment and reconstruct face images in the frontal pose and with neutral illumination (we call it the canonical view) from face photos in the wild. This project will also study recognition algorithm for low resolution and blurred face images. Given the importance of data size in the training of deep models, we also propose to construct a large face database to overcome various limitations of existing datasets. The outcome of this research will advance the technologies of face recognition in complex and uncontrolled environments.

伴随着移动终端摄像头的普及和移动互联网的发展，互联网上照片视频数量快速增加。如何高效鲁棒的识别网络图像中的人脸已经成为一个亟待解决的问题。网络照片和视频往往在不可控的自然条件下拍摄，姿势、光线、表情和化妆等变化复杂，且人脸图像的分辨率可能很低，这对传统的人脸识别技术带来巨大的挑战。本项目以机器学习为主要工具，从底层特征的学习、复杂条件下人脸的比对和识别算法等多方面展开研究。我们将建立一个大的人脸数据库对算法进行验证。本项目对于推动人脸识别技术在复杂、不可控环境下的广泛应用有着重要的意义。

本项目以深度学习为主要工具,针对实际场景中光照、遮挡、姿态变化、尺度等对人像检测和识别带来的挑战，研究开发新型深度人脸检测和识别模型。主要成果包括：在人脸检测方面，提出级联人脸检测方法，通过级联三个深度神经网络，并联合人脸和对齐（关键点检测）两个任务，融合人体区域，实现高精度人脸检测。在人脸识别领域，针对非受控环境下复杂姿态和光照所带来的变化,利用同一个人人脸的聚类效应提出中心损失函数;针对训练数据中不同类别的不平衡问题,提出极差损失函数;并建立一个大的人脸数据库进行验证。项目执行期间共发表SCI/EI论文25篇，其中在视觉领域著名期刊T-IP上发表论文1篇,在计算机视觉三大顶级会议CVPR、ECCV、ICCV上发表论文15篇，申请发明专利6项。