Superposed Linear Representation in Reproducing Kernel Hilbert Space for Image Classification

The non-parametric subspace classification to obtain excellent classification performance mainly depend on directly using the training samples as the dictionary of each class. The most typical classification algorithm is the collaborative representation-based classification. Recently, a superposed linear representation classifier that expresses recognition problems by presenting test images as the superposition of centroid and shared intra-class differences reveals that the superior performance of the collaborative representation depends to a large extent on the significant enough class separability of the controllable face datasets. However, the superposed linear representation classifier does not take account of the nonlinear features hidden in the visual features. In the paper, we propose to utilize the kernel technique to extend the superposed linear representation classifier method. our proposed method tested on several databases, such as handwritten digital recognition, image classification, and fine-grained recognition, and obtains satisfactory classification performance.