Supergaussian data denoising by semi-ICA estimation

In this paper, a novel nonlinear denoising algorithm, semi-ICA estimation, is developed to remove noise with arbitrary distribution from correlated supergaussian signals. Semi-ICA estimation consists of an orthogonal linear transformation and a Bayesian based nonlinear estimation function. The linear transformation maximizes the dissimilarities between original signals and the noise in the sense of nongaussianity and achieves the independency among components of signals. The component-wise Bayesian based nonlinear estimation is to estimate the components of the clean signal in the transformed domain. Based on these two objectives, a recursive algorithm, maximizing the difference of negentropy between signal and noise and an estimation function, minimizing mean square error, is derived. Results of the proposed technique are compared with the results of sparse code shrinkage (SCS) and Wiener filter for random signal data.