Blind source separation using support vector machines

Blind source separation (BSS) can be defined as the recovering of source signals from their mixture signals without any prior knowledge of the source and mixing environment. The type of the mixing environment may be linear or nonlinear and this determines the type of the BSS problem. Independent component analysis (ICA) method solves linear BSS by finding a representation for a multivariate data that minimizes the statistical dependency between signal components, but ICA can not find solution for nonlinear BSS. The kernelbased methods have been implemented to solve the nonlinear BSS problems. The temporal predictability method that is based on the measure of temporal predictability solves the linear BSS problems, but it can not achieve the nonlinear BSS. The kernelized temporal predictability method is based on applying the kernel methods in support vector machines (SVMs) to the temporal predictability method. In this thesis, the linear and nonlinear BSS applications have been simulated in Matlab numeric analysis software package by using the linear and kernelized temporal predictability methods, respectively. In addition to sound separation, the algorithms has been extended for the image separation problem newly in this study. The statistical performance of the algorithms have been simulated using different algorithm parameters like kernel degree, kernel type, image mass size and nonlinear mixing function.