Abstract: Multi-kernel SVDD is proposed on the basis of SVDD. Due to the disadvantage of having too simple a kernel formation, the SVDD is expanded from a single Gaussian kernel to a linear combination of multi-Gaussian kernels. The resulting multi-kernel SVDD could be expressed as an SDP problem, so it could be solved with the global optimal solution. The proposed method employs a more complicated kernel formation, which can describe the distributing boundary of training examples in high-dimension feature space more flexibly, thus leading to a much higher recognition rate and a much lower false rate than the SVDD.

Key words: recognition, rejection, multi-kernel support vector domain description, semidefinite programming, global optimal solution