Quasiconformal-matrix-based multikernels learning for sensory data classification

Sensory data classification is a crucial task in many information processing. The vector data or matrix data are the most common data format, especially matrix data, for example, image data. We propose Gaussian kernel constructing method adaptive to the distribution of the input data for classification. In the current matrix based feature extraction, the matrix image is transformed to the vector, and the procedure of transforming will increase the ability of a large saving space and computing. And secondly the different geometrical structures under the different kernel function will bring the different class discrimination of the input data in the feature space. The performance of kernel will be increased under the adaptive choosing the parameters of kernel. We present the matrix Gaussian kernel, Quasiconformal matrix Gaussian kernel, and Quasiconformal matrix multikernels, and we implement the three kinds of kernels on different classifiers on ORL and Yale image databases. The experimental results show that the proposed kernels perform better than the traditional kernel functions.