The Improved Natural Gradient Algorithm And Its Application For The Voice And Image Processing

The natural gradient algorithm is a common algorithm for blind source signal processing applications, especially in the blind source separation and independent component analysis. As one of the core of the algorithm, this algorithm will be more and more attention. Because of its convergence speed, the separation effect, etc., which highlighting its a good prospect, and have a wide range of applications in many ways.This paper focuses on the shortcomings of conventional natural gradient algorithm, as well as the application of the voice and image to do the following work:Summary and analysis of the natural gradient algorithm for blind source separation and its related theories,and focus on the analysis and comparison of the natural gradient algorithm on a fixed step size and variable step size. There are some drawbacks:fixed step natural gradient algorithm for step demanding a direct impact on the convergence speed and stability; its step change of the variable step size natural gradient algorithm can not demand change. Although it can improve part performance, the convergence rate and the separation effect is not ideal under complicated conditions and in the application process.This article describes the natural gradient algorithm of information theory and the basics of the higher-order statistics and blind source separation and independent component analysis of several criteria and algorithms. Focused on the basic principle of three most commonly used algorithm:the minimization of mutual information (MMI), the information maximization (Infomax) and maximum likelihood estimation (MLE). Analysis the principles and procedures of the natural gradient algorithm, and introduce the momentum factor in the neural network and the concept of the degree of separation of signal separation, which be introduced into the natural gradient algorithm in the step natural gradient algorithm and improve it. The improved algorithm standard by the degree of separation, change its step size through adaptive changes in momentum factor adaptively, and thus able to improve blind source separation performance of the signal. Give an example of A set of linear instantaneous mixed-signal, it successful implement the blind source separation, and proved its superiority through contrast by the algorithm improved before. The improved algorithm is also used in speech separation and image denoising. Through a set of instances of speech separation and denoising of images with noise from the image,after the separation, as well as the three evaluation to the qualitative and quantitative analysis, it illustrates the improved algorithm has a unique advantage.