Kernelized discriminative–collaborative representation-based approach for pattern classification

• Published in: Computers & electrical engineering 2022-10, Vol.103, p.108342, Article 108342
• Main Authors: Wang, Shuang-xi, Ge, Hong-wei, Gou, Jian-ping, Ou, Wei-hua, Yin, He-feng, Liu, Guo-Qing, Halimu, Yeerjiang
• Format: Article
• Language: English
• Subjects: Discriminative–collaborative representation; Nonlinear representation; Pattern classification; Uncorrelated sparsity constraint
• ISSN: 0045-7906; 1879-0755
• DOI: 10.1016/j.compeleceng.2022.108342

In representation-based classification methods, Gaussian function is adopted for realizing the nonlinear representation of samples. However, there are some problems. The definition of Gaussian kernel includes exponential operation, which causes the parameters to be sensitive and consumes exponential computational time. In addition, the performance of pattern classification may be dropped with noise (outliers). To solve these problems, a new Euclidean Kernel is designed to replace Gaussian Kernel. Moreover, a new weight and an uncorrelated sparsity constraint are introduced to enhance within-class representation consistency and between-class representation discrepancy. Then, we propose a kernelized discriminative–collaborative representation-based approach for pattern classification by incorporating the cooperative and competitive representation. The proposed model fully excavates the intrinsic factors of the samples and presents promising performance and better interpretability. Finally, extensive experiments are conducted in comparison with other popular methods on different types of databases, and these experiments verify that the designed approach performs better than its competitors. [Display omitted] •A new Euclidean Kernel and a new weight are constructed to capture the intrinsic structural features and nonlinear attributes among samples.•An uncorrelated sparsity constraint is designed to reduce the correlation among different-label samples and generate a sparse solution.•We propose a kernelized discriminative–collaborative representation-based approach to transform the samples from low-dimensional space into the discriminative nonlinear one.