基于深度学习的人脸识别关键技术研究

In the age of big data in vision, face recognition has broad application prospects and important theoretical significance. To effectively exploit large-scale face data for improving the face recognition performance, the project intends to investigate the four key techniques of face recognition based on the deep learning theory, including face detection, face clustering, face normalization and face matching, and proposes correlation filter based region proposal generation, robust statistics based face clustering, multi-task deep learning based face normalization, transfer deep learning based metric learning, respectively, so that the face recognition difficulties caused by various factors, such as pose, expression, illumination and noise, can be overcome. The project team has a good foundation for the face recognition research. This project will further promote the development and progress for the theory and technology of face recognition.

视觉大数据时代下的人脸识别研究具有广泛的市场应用前景和重要的理论指导意义。为了有效地利用人脸大数据来提高人脸识别的性能，本项目以深度学习理论为基础，针对人脸识别中的人脸检测、人脸聚类、人脸归一化、人脸比对四个关键技术进行深入研究，分别提出基于相关滤波器的区域建议生成算法、基于鲁棒统计学的人脸聚类算法、基于多任务深度学习的人脸归一化算法，以及基于迁移深度学习的度量学习算法，从而克服姿态、表情、光照、噪声等各种复杂因素给人脸识别带来的诸多技术瓶颈问题。本项目组具有良好的人脸识别研究工作基础。对本项目的研究将进一步推动人脸识别理论和技术的发展和进步。

视觉大数据时代下的人脸识别研究具有广泛的市场应用价值和重要的理论指导意义。为了有效地利用人脸大数据来提高人脸识别的性能，本项目以深度学习理论为基础，针对复杂环境下的人脸识别问题，从人脸检测、人脸聚类、人脸归一化、人脸比对四个关键技术展开深入研究，建立起一套有效的人脸检测、聚类、归一化和对比方法和系统方案，从而克服姿态、表情、光照、噪声等各种复杂因素给人脸识别带来的诸多技术瓶颈问题。项目组成员共发表与项目研究内容相关学术论文60篇，其中SCI检索论文31篇，EI检索论文27篇，JCR一区论文4篇，JCR二区论文18篇，中国计算机学会（简称CCF）推荐的A类期刊论文2篇，CCF推荐A类会议2篇。申请发明专利10项，其中已授权发明专利9项。