Feature level and decision level fusion in kernel Sparse Representation based Classifiers

Abstract

Sparse Representation based Classifier (SRC) is widely used in a variety of pattern recognition and machine learning tasks. The kernelized version of the classifier (Kernel SRC) attempts to remove the SRC limitations by a nonlinear mapping of the data through kernel functions. However, the performance of such a method strongly relies on the choice of the kernel function. In this paper, we firstly create composite kernels by combining a set of basis kernels. This process can be seen as a feature level fusion of the single kernel classifiers. In this step, the basis kernels are linearly combined using different rules. Within the framework of the SRC, in the next step, the reconstruction errors of the resulted classifiers are combined in decision level in order to make the final decision. Our experimental results on different datasets show the merit of the proposed method.