Research On Algorithm And Applications Of Independent Component Analysis With Reference

During the past decade, independent component analysis (ICA), also called blind source separation has become an active branch of signal processing. It can recover independent source signals from their observed mixtures without knowing the distribution type of the source signals and the mixing coefficients. Owing to the specific advantage of weak request for prior information, ICA has been applied to many fields such as wireless communication, biomedical signals processing, images and speech processing.Traditional ICA aims to recover all independent source signals. However, many real situations only need one source signal or several source signals of interest. For example, it is the target speech instead of noises that is desired for speech enhancement. Therefore, independent component analysis with reference (ICA-R) was proposed by incorporating some prior information about the sources into the conventional ICA to extract the desired sources under some measurements. Compared with traditional ICA, ICA-R has fast convergency speed by using a little prior information. It has been efficiently applied to fMRI (functional Magnetic Resonance Imaging) data analysis.This thesis contributes to ICA-R in the following four aspects. (1) We apply ICA-R to speech enhancement by properly constructing the reference signals of target speech signal. We analysis the speech characteristics and differences between the speech signal and many noise signals. The pitch period and the envelopes of the power spectral of speech are utilized to construct two kinds of reference signals. The performances of ICA-R under closeness measurements of mean square error and correlation are compared for different noisy mixtures of speech. (2) We also apply ICA-R to extract fetal electrocardiogram (FECG) from measurements taken from pregnant woman by constructing the reference signal of FECG. This is also useful for other ICA-R applications such as EEG/MEG. (3) During the applications of ICA-R, we find that its contrast function is complex, the complicated matrix computation is involved, and the stability and convergence are thus affected. To solve these problems, we give an improved ICA-R by pre-whitening the observed signals and by normalizing the weight vector. (4) Since the original ICA-R algorithm is sensitive to the initial value and its learning rate parameter needs to be defined by users, we present a fixed-point ICA-R algorithm to eliminate the limitation. It is easy to use and more reliable.Extensive computer simulations and performance analyses demonstrate the efficiency of all our work.The research on ICA-R in this thesis is helpful for proposing novel semi-blind algorithms. Besides, it has more promising applications in extrating source signals of interest than the traditional ICA.