基于稀疏子空间聚类的跨域人脸迁移学习方法  

Cross-domain face transfer learning based on sparse subspace clustering

作　　者：朱俊勇[1,2] 逯峰[2] 

机构地区：[1]中山大学数据科学与计算机学院,广东广州510006 [2]广州市刑事科学技术研究所,广东广州510050

出　　处：《中山大学学报（自然科学版）》2016年第5期1-7,共7页Acta Scientiarum Naturalium Universitatis Sunyatseni

基　　金：中国博士后科学基金资助项目(2015M582469);国家自然科学基金资助项目(61573387)

摘　　要：人脸识别的效果很大程度上依赖于已标定的训练数据的规模,当训练样本严重不足时类内及类间分布的估计将会出现严重偏差。考虑到人工标定的成本过高,如果能对与目标问题相关的一些已有数据加以利用,以此来取代人工标定数据或减少人工标定的数据量,将为训练样本不足的人脸识别问题提供一套可行的解决方案。为此,拟针对这一问题发展出一种基于稀疏子空间聚类和鲁棒主成分分析的人脸迁移学习方法,在辅助数据满足多线性子空间假设下,能从无类标的异源辅助数据中实现信息迁移,挖掘对目标分类问题有益的成分。The quality of a face recognition system heavily depends on the amount of labeled training data. Bias would probably exist in both in-class and between-class scatter when there is few labeled data. Considering the cost of manual labeling is too high, an alternative choice is to make use of existing data which is related to the objective problem. In this way, it is able to alleviate the dependence of manual labeling via exploring numerous related data, offering a feasible solution to the case of lacking sufficient labeled training samples. To this end, a face transfer learning approach based on the sparse subspace clustering and robust principal component analysis is proposed, which allows employing unlabeled source data under the multi subspace assumption and mining useful information for the objective problem.

关 键 词：稀疏子空间聚类 低秩矩阵分解 鲁棒主成分分析 跨域人脸迁移学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]