Independent Component Analysis And Its Application To Feature Extraction

Independent component analysis(ICA)consists of recovering the potentially independent source signals by analyzing the higher-order statistical correlation of the multidimensional data without any prior knowledge of the source signals and mixing matrix.Nowadays,the ICA technique has a lot of interesting applications in various fields,such as biomedical signal,image processing and speech signal processing.This paper mainly discusses the basic theories,the algorithms of ICA and its applications in the natural image feature extraction.Furthermore,a new ICA algorithm is proposed by optimizing the maximum likelihood ICA criterion with the weighted orthogonal constrain natural gradient.The main work included in this paper are summarized below:(1)The ICA development,research status and its special applications are introduced briefly.Meanwhile,the ICA theories are summarized systematically.The ordinary process of solving the ICA algorithm mainly included data preprocessing,objective function and its optimization.Finally,some kinds of classical ICA algorithm are presented in detail.(2)The ICA algorithm is applied to the natural image feature extraction.Firstly,we show the application of the principle component analysis to the image feature extraction,and analyze its deficiencies.And then,we introduce the higher-order statistics based ICA algorithm into the extraction of the image features.The ICA algorithm uses fully the higherorder statistics of the image data,a set of independent feature vectors are obtained.Computer simulation results show that the Infomax algorithm and Fast ICA algorithm can successfully extract the basis vectors.The most of them are localized in space,frequency,and orientation,which describe the edge features of the natural images well,and the corresponding coefficients obey stronger super-Gaussian distribution.(3)A new ICA natural gradient algorithm based the maximum likelihood is proposed.The ICA algorithms based on the pre-whitened data not only lack the equivalence property,the propagation of pre-whitened error will deteriorate the accuracy of the estimate of ICA.To overcome the above weakness,we establish the weighted orthogonal constrain maximum likelihood ICA criterion.Then,solve it by the Lagrange method and natural gradient,a new ICA algorithm is proposed.Simulation on man-made signals show that the new algorithm can separate and recover the sources signals efficiently,and has a better convergence speed than other ICA algorithms.(4)Furthermore,the new algorithm is applied to the image feature extraction.The experiment show that it can extract the independent features successfully.On the other hand,we also present some additional extensions of basic ICA model by relaxing the assumption of the independence,which is used to obtain the nonlinear features.Simulation results show the features in same independent subspace have the similar frequency,orientation and location.