Weighted Discriminative Dictionary Learning based on Low-rank Representation

2016 International Conference on Communication, Image and Signal Processing

物理学;无线电电子学;计算机科学

Chang, Heyou^1 ; Zheng, Hao^1

Key Laboratory of Trusted Cloud Computing and Big Data Analysis, School of Information Engineering, Nanjing XiaoZhuang University, Nanjing, China^1

关键词: Classification performance;    Discriminative dictionaries;    Low-rank representations;    Regularization terms;    Representation-matrices;    Semantic dictionaries;    State-of-the-art methods;    Training and testing;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/787/1/012020/pdf
DOI  :  10.1088/1742-6596/787/1/012020

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

Low-rank representation has been widely used in the field of pattern classification, especially when both training and testing images are corrupted with large noise. Dictionary plays an important role in low-rank representation. With respect to the semantic dictionary, the optimal representation matrix should be block-diagonal. However, traditional low-rank representation based dictionary learning methods cannot effectively exploit the discriminative information between data and dictionary. To address this problem, this paper proposed weighted discriminative dictionary learning based on low-rank representation, where a weighted representation regularization term is constructed. The regularization associates label information of both training samples and dictionary atoms, and encourages to generate a discriminative representation with class-wise block-diagonal structure, which can further improve the classification performance where both training and testing images are corrupted with large noise. Experimental results demonstrate advantages of the proposed method over the state-of-the-art methods.