Research And Application Of Stochastic Gradient Variational Bayesian Learning Model With G~0 Distribution

Over the years,with the development of remote sensing technology,SAR images have been widely used in many fields.The basis of SAR image processing is SAR image segmentation.However,due to the special imaging of SAR images,the extremely heterogeneous regions are always difficult.At present,there are ways to use artificially proposed features for segmentation,but this requires professional background knowledge.Its cost and efficiency are relatively high.In machine learning,the Bayesian learning model can be inferred using prior knowledge and can adaptively learn the characteristics of the data.This model can be used to learn the characteristics of extremely inhomogeneous regions.Therefore,in view of these situations,a SAR image segmentation method that is based onG~0 distributed stochastic gradient Bayesian learning model is proposed.The main innovations of this method are as follows:(1)A stochastic gradient Bayesian model that is based onG~0 distribution is proposed.Bayesian networks need to carry out probabilistic assumptions on hidden variables and approximate posterior probabilities.The Gaussian distributions used in optical images do not fit well with SAR image data,so the features of SAR images cannot be effectively learned.TheG~0 distribution can well characterize the statistical features of SAR images.Therefore,this thesis combines the Bayesian network andG~0 distribution.The assumptions of the probability distribution in the network isG~0 distribution.Then this thesis constructs a variational Bayesian learning model based onG~0 distribution,and then in both n(28)1 and n?1 cases,this thesis deduces the analytic form of the lower bound of the variation in the corresponding learning model.Thereby it provides a model framework for the subsequent learning of the characteristics of extremely inhomogeneous regions in the hybrid pixel subspace of the SAR image.(2)A SAR image segmentation method which is based onG~0 distributed stochastic variational Bayesian model is proposed.For the hybrid pixel subspace in the SAR image semantic space,the features of the extremely inhomogeneous region are learned using the G~0 distributed variational Bayesian learning model designed in(1),and then the hybrid pixel subspace is segmented using the learned features.The first thing is to construct a variational Bayesian learning model based onG~0 distribution for each extremely inhomogeneous region,and according to the probability density function obtained for each region,a corresponding set of matrices can be obtained.Then these matrices can be as initialization for each region.The trained weights can be as the characteristics of the corresponding area.Next,according to the idea of the bag of feature,the features learned in all areas are connected together to form a codebook,and the projection of each area to it is calculated,and the features representing the corresponding areas are obtained by the method of maximum convergence.Finally,the classification of these features is obtained through hierarchical clustering,that is,the classification of the corresponding extremely inhomogeneous area.That is the segmentation result of the hybrid pixel subspace.(3)For the network proposed in(1),two parameters in the network are analyzed.One is the number of input neurons,which is the input image block size.The other is the number of hidden neurons.First of all,for the parameters to be analyzed,the network in the(1)is trained with different values,and then the clustering method is used to analyze the number of feature clusters for the learned features.The number of clusters reflects the difference between the features.That is to say,the greater the number of clusters,the more features that are learned.