Blind Source Separation Algorithm Research Based On Constraint And Multi-task Learning

With the development of information technology,image processing technology has made significant progress,and been widely applied in the medical,remote sensing,military,daily life,and many other fields.Blind source separation problem is a classical issue in signal and image processing.It aims to recover unobserved source signals from their observed mixtures without any information of the mixing system.In recent decades,blind source separation technology has received wide attention of scholars both at home and abroad and widely used in the speech signal processing,medical image processing,hyperspectral remote sensing image processing,and other fields.In this paper,we study some blind source separation algorithm based on statistics and its related theory.The main contents are as follows:1.Firstly,we analyze some classic blind source separation algorithms including maximizing non-Gaussian independent component analysis and ICA based on maximum likelihood estimation.Independent component analysis(ICA)is an important method in blind source separation methods,many algorithms are based on this method.Nonnegative matrix factorization(NMF)algorithm aims to decompose the original matrix into two nonnegative matrixes.NMF has also obtained good effect in dimension reduction,classification problems and so on.2.NMF usually falls into local optimal solution.It is the main issue that NMF cannot obtain the global optimal solution.In this paper,we proposed KL divergence and sparse constrained non-negative matrix factorization algorithm(KLSNMF).KLSNMF can fully use the structure information of the dataset and get the global optimal solution.The effectiveness of proposed method is shown in the experiment part.3.Most of existing BSS methods independently decomposes each mixture signal without a thorough investigation on the decompositions of others.The underlying relationships between mixture signals,which are widespread in practice,are therefore discarded.In this paper,we regard the decomposition of each mixture signal as a task.Thus separating source signals simultaneously for multiple mixture signals naturally leads to a multi-task learning problem.The multi-task decomposition process can make full use of the underlying connections between tasks to improve the performance of unmixing algorithm.The proposed method obtained great decomposition result in simulated dataset,natural images and hyperspectral remote sensing dataset.The research works in this paper are closely related with existing blind source separation researches.This paper has certain theoretical and application values to push forward the theories and algorithm of blind source separation research.