海量视频数据中基于异构并行系统的人脸识别技术

With the explosive growth of video data volume of human society, there are low efficiency and low accuracy problems of face recognition for massive video data now, and parallel computing is one of the optional effective ways for face recognition in the massive video data processing. Parallel implementation of the issues facing the whole process of video-based face recognition including research face target detection and tracking, facial feature extraction and face target recognition algorithm are to calculate the amount of a typical high-performance applications that perform commence analysis and discussion, and to significantly improve the efficiency of face recognition massive video data. The study includes: (1)Face recognition algorithm parallelization based on the small dictionary learning classifier; (2)Video face descriptive model based on the manifolds; (3)Parallel AdaBoost algorithm for the massive video data of human face target detection; (4)Pedestrians face target tracking algorithm based on MeanShift algorithm and Kalman filter; (5)PCA face feature extraction based on heterogeneous parallel architecture block. This study will not only provides new ideas for the design of face recognition algorithms, but also lies the foundation for the improvement of the efficiency of parallel applications in face recognition.

随着人类社会的视频数据量爆炸性增长, 目前针对海量视频数据的人脸识别存在效率低和准确率低等问题, 并行计算是海量视频数据中高效进行人脸识别处理的可选途径之一. 本项目面向海量视频数据中人脸分析与识别全过程的并行化实现, 研究人脸目标检测和跟踪、基于稀疏表示的人脸特征提取与识别算法, 以显著提高海量视频数据中的人脸识别效率. 研究内容为面向海量视频数据的: (1)基于小容量字典学习的人脸分类算法及其并行计算; (2)基于流形的视频人脸描述模型建立; (3)人脸目标检测的并行AdaBoost算法; (4)基于MeanShift算法和Kalman滤波器的人脸目标跟踪算法; (5)基于异构体系架构下的并行分块PCA人脸特征提取算法. 本项目的研究将不仅为人脸识别算法的设计提供新的思路, 而且为人脸识别并行应用效率的提高和为大规模人群背景下的人脸分析与识别的实际应用奠定基础.

本项目属于人脸识别领域，针对基于海量视频中人脸识别的高性能并行计算，探索性地研究异构体系架构下的人脸识别并行化方式，研究内容已经按计划完成。. 项目组针对研究目标，采用加权稀疏表示方法建立了面向海量视频数据中的并行人脸识别计算模型，利用异构体系架构下的并行分块进行PCA人脸特征提取，基于KSVD的稀疏表示字典学习算法的并行计算方法，缩短了人脸识别的计算时间；同时建立多核异构并行系统的人脸识别实验平台，设计了适合典型高性能计算应用的并行人脸识别软件系统原型。相关研究成果正在寻求推广应用，已经与合作单位进行应用项目的开发。. 本项目共发表论文9篇，其中7篇为 SCI收录论文，2篇进入EI检索。申请专利10项，申请软件著作权4项。培养博士3人，硕士12人。