Polarimetric SAR Image Classification Based On Sparse Learning And Deep Directionlet Network

From the beginning of the study of polarimetric SAR images up to now,nearly thirty years have passed,many methods of classic polarimetric SAR image classification have been widely used in various fields of real life.In the study of polarimetric SAR image,it can be said that researchers have achieved very good results.Nowadays,the research results of polarimetric SAR have already begun to be civil,which requires higher resolution of polarimetric SAR images.Although the technology has been improved a lot in obtaining polarimetric SAR data,the polarimetric SAR image processing technology still has much room for further development.In this thesis,we mainly study polarimetric SAR image classification method based on sparse learning and deep Directionlet network.The deep neural network continuously abstract the low-level features of SAR and highly approximate the complex functions,and then extract the deep features in the polarimetric SAR images.Sparse learning can reduce the redundancy of polarimetric SAR data,which can extract feature information efficiently.In this paper,three different polarimetric SAR image classification methods are proposed.In this paper,the features of polarimetric SAR images are extracted by sparse learning and deep Directionlet network,and then the classification results are obtained.The main work is summarized as follows:A polarimetric SAR image classification method based on multilayer SVM and super pixel sparse coding is proposed.This method firstly selects a certain proportion of training samples,uses the SVM algorithm to classify,and then according to the classification results of the SVM algorithm,all the sample points are sampled.For the important sample points,on the one hand,it is used for the next SVM classification,and so on,so as to form a multi-layer SVM.In addition,it is also used in the training of the dictionary and the classification of the super pixel sparse coding method.In the classification results of multilayer SVM,classification results of similar objects is bad,but the boundary of classification results is good and in the results of super pixel sparse representation classification,classification results of similar features is good,but the boundary of classification effect is poor.Therefore,according to the classification results of the two methods,we can get better classification results by using voting mechanism.A polarimetric SAR image classification method based on deep Directionlet neural network is proposed.The method is based on the convolution neural network framework,and then the filter banks of the roll up layer are obtained through the Gauss wavelet basis function,and then each filter is rotated to different angles to obtain a new filter bank.The new filter bank is used to segment the image blocks,and then the deep level features are extracted by the next sampling layer.Using the new filter bank CNN not only can obtain the wavelet information,but also can get different direction information,which is helpful to extract more abundant information,and further improve the classification accuracy.A polarimetric SAR image classification method based on multi-scale deep Directionlet neural network is proposed.Because the size of the traditional filter can only be set,which limits the flexibility of the parameters,the global and local features can not be extracted at the same time.Therefore,using a variety of different sizes of the filter method in the convolution of the same image block,which can extract local features of images,but also can extract global features,the formation of multiple pathways in fully connected layer multi channel merge,called multi-scale depth Directionlet network.The experimental results show that the classification accuracy of the proposed method is better than that of the deep Directionlet network.