Nonlinear Noise Filtering with Support Vector Regression

This paper introduces the novel application of support vector machine for filtering of time-series signal corrupted by Gaussian and non-Gaussian noise. In real world, the case of non-Gaussian noise (e.g. impulse noise, signal dependent noise) is very common. The optimal Wiener filter, which is a linear approach, can yield good results to Gaussian white noise, but performs poorly in case of non-Gaussian noise. Considering the noise filtering problem as a mapping of noisy signal to the corresponding noise free signal, we utilize the support vector regression (SVR) tool to discover the dependency so as to implement the noise filter. Comparing with the Wiener filter, SVR performs better in case of non-Gaussian. In this paper, we generate the original signal by an AR model, and then corrupt them with Gaussian and impulse noise respectively. The performance of mean squared error (MSE) is compared with wiener filter and multiple layer perceptron (MLP).