包含非连通组的动态组结构稀疏人脸识别方法  

Face Recognition Method Using Non-Connected Dynamic Group Sparse

作　　者：蔡体健[1,2] 樊晓平[1,3] 徐君[2] 谢昕[2] 郑轶[2] 

机构地区：[1]中南大学信息科学与工程学院,长沙410075 [2]华东交通大学信息工程学院,南昌330013 [3]湖南财政经济学院"财经大数据资产开发与利用"湖南省2011协同创新中心,长沙410205

出　　处：《计算机辅助设计与图形学学报》2016年第4期590-597,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家科技支撑项目(2012BAH08B01);江西省科技厅工业支撑重点项目(20151BBE50055);江西省自然科学基金(20142BAB207007);教育部人文社会科学研究青年基金(14YJCZH172)资助

摘　　要：针对稀疏分类模型中存在着非连通的组结构,为提高模型的表示能力,提出一种非连通的动态组结构稀疏人脸识别方法.该方法采用组合方式搜索所有可能的基块,包括非连通基块,通过基块的联合构造动态组结构;将高维数据按类别分块,在小的分块内进行组合搜索,避免了组合爆炸;采用编码复杂度来衡量数据的结构稀疏度,给出各种结构的编码复杂度计算方法;基于结构贪婪算法实现非连通的动态组结构稀疏重构.最后在AR,Extended Yale B和CMU-PIE人脸库上进行实验,验证了文中方法的有效性及稳定性.Sparse representation-based classification model contains non-connected group structure. The proposed face recognition method utilizes non-connected dynamic group sparse to improve representation ability of model. The method employed combinatorial search to obtain all possible base-blocks, including non-con-nected base-blocks, and constructed dynamic group structure by union of the base-blocks. Meanwhile the method divided high dimensional data into blocks according to human category, and performed combinato-rial search within block in order to avoid combinatorial explosion. Furthermore the method adopted coding complexity to measure structural sparsity of coefficients, and analyzed calculation methods of coding com-plexity for several sparse structures; then achieved sparse reconstruction of non-connected dynamic group structure based on structured greedy algorithms. Finally, the face recognition experiments on the AR, Ex-tended YaleB and CMU-PIE databases demonstrate the effectiveness and stability of the proposed method.

关 键 词：稀疏表示分类 结构稀疏 结构贪婪算法 组合搜索 连通性 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]