| During the past decade, independent component analysis (ICA), also called blind source separation has become an active branch of digital signal processing. ICA can recover independent source signals from their observed mixtures without knowing the distribution of the source signals and the mixing coefficients. Owing to its specific advantage of weak request for prior information, ICA has been applied to many fields such as wireless communication, biomedical signals processing, and image and speech processing.Traditional ICA algorithms usually have some problems such as ambiguous outputs and low estimation efficiency. In the mean time, some piror information is available about source signals and the mixing parameters. As such, some prior information is directly utilized in BSS, which is just called semi-blind ICA. Among various algorithms, ICA with reference (ICA-R) is a typical one with determined outputs and efficient estimation. However, ICA-R focuses only on real-valued signals though the complex signals are also very practical, e.g., the communication signals and the speech signals in the frequency domain.This thesis contributes to comlex-valued ICA-R in the following three aspects. (1) By analyzing its characteristics, the complex ICA model is converted into a real-valued ICA. ICA-R is then applied to the real part or the imaginary part to give the corresponding estimations of the desired signal. (2) In the framework of constrained ICA, the complex ICA-R is formulated as maximizing the contrast function of a blind complex fastICA algorithm under an inequality constraint corresponding to the magnitude information. a fixed-point algorithm and a Newton algorithm are then derived, respectively. (3) A new band-pass filter method for constructing reference from noisy speech is proposed. Speech enhancement is then performed in the time domain and frequency domain for comparison. Computer simulations and performance analyses demonstrate., the efficacy of the above-mentioned methods. |
PDF Full Text Request