An improved redundant dictionary based on sparse representation for face recognition

详细信息    查看全文

• 作者：Fanrong Meng ; Zhenyu Tang ; Zhixiao Wang
• 关键词：Sparse representation ; Face recognition ; Redundant dictionary ; Feature fusion ; LDA (linear discriminate analysis)
• 刊名：Multimedia Tools and Applications
• 出版年：2017
• 出版时间：January 2017
• 年：2017
• 卷：76
• 期：1
• 页码：895-912
• 全文大小：
• 刊物类别：Computer Science
• 刊物主题：Multimedia Information Systems; Computer Communication Networks; Data Structures, Cryptology and Information Theory; Special Purpose and Application-Based Systems;
• 出版者：Springer US
• ISSN：1573-7721
• 卷排序：76

文摘

In recent years sparse representation has been widely used for face recognition and achieved good results. Most sparse representation methods need a redundant dictionary to solve sparse coefficients. And the number of atoms must be much larger than the dimension of atoms in the dictionary. So the design of redundant dictionary is very important for improving the performance of sparse representation methods. By experiments we find that feature fusion (LBP, Gabor, Hog, and raw pixels) after PCA can remain a high recognition rate, which means the feature fusion can represent faces well with a low dimension. So we can use the dictionary based on feature fusion to solve small sample size problem in LDA without losing useful information. And LDA can increase between-class scatter and decrease within-class scatter while reducing the dimensionality, which can build a better structure for redundant dictionary. Based on above we propose a linear discriminative redundant dictionary based on feature fusion to improve the performance of face sparse representation methods, namely, LDRD. Firstly, extract and concatenate a standard set of features (LBP, Gabor, Hog, and raw pixels) to form a feature vector as the atoms, then introduce LDA to rebuild the dictionary of atoms, to reduce dimensionality and enhance the discriminative ability of the dictionary. We compare LDRD with the dictionary based on downsampling and feature fusion for SRC, CRC_RLS and LASRC. The extensive experiments demonstrate that the proposed dictionary has better recognition rate and operating efficiency, while it can easily reject distractor faces.