Blind Source Separation with Image De-Noising using SVD Regularization

Abstract

The aim of this report is to provide a comprehensive overview of neural networks for blind source separation scheme along with mathematical foundation. Blind source separation (BSS) is scheme of separating source signals from a set of mixed signals without having any information or with very little information about the source signals or the mixing process. So BSS usually assumed that source signals count is known as priori. Typically it should be equivalent to the number of sensors and mixtures. The analysis of BSS using neural networks which are separation rule with prewhitening, Global rule, Local Rule for detecting and extract the presence of the useful source signal from mixed signal, along with the different Kurtosis conditions are also explained. The method is proposed for separating the images from mixtures. The principle of proposed method includes BSS scheme followed by a SVD regularization procedure. The SVD has potential to smooth the source image through regularizations. The proposed scheme not only reduces the noise but enhance the quality of source images also. The problem of less sensor count then sources are try to be address out in the simulated process, in which compression of data is done first in prewhitening stage then in separation stage where there are more mixtures then original images 􀝐 > 􀝎, but the results are good in prewhitening than separation stage since compression in separation stage enhanced the noise. Also the redundancy elimination is describe for both noise free and noisy environment where sources are fever then mixtures then single layer global rule is applied. The performance of proposed scheme is compared with existing BSS schemes based on parameters such as PSNR, MSE, SSIM and EI.