SAR Image Classification Based On Transform Domain Features

In recent years,SAR image classification has received widespread attention.We hope to use a little part of the image to classify the whole image with a larger size.However,the traditional image classification methods can not be applied in SAR image directly due to the existence of speckle noise.During the development of sparse representation,the usefulness of signal sparse representation has been realized.But in the past,people have been focusing on the study of sparse synthesis model.Large number of literature paper learning an overcomplete dictionary to provide sparse synthesis coefficients and use them as the feature of image.Another sparse representation model—cosparse analysis model has been ignored.The advantages of this model have gradually been emphasized recently.And trying to apply it into the image denoising?recovery and other fields of image.Here are the main elements of this paper:Firstly,we modify the cospa rse operator learning method which have been used in image denoising,and make sure that it can be applied in the image classification correctly.We select the appropriate size pixel block with the center of every pixel as the initial signal,and use previously modified operator learning method to update the analysis operator and the original signal alternately based on the analysis sparse model.Then,the Augmented Lagrangian method is applied to solve the cosparse coefficients.And the obtained coefficients as the features of the classification samples will be put into the SVM classifier for classification.Secondly,considering that the transform domain features are not enough to distinguish the details of the image,we decided to combine the texture feature and the transform domain feature.We use sparse autoencoder to extract the output features of the cosparse coefficients,and mix the features and gray value of initial pixel block as the final classification feature.We also do our experiment based on SVM classifier.We test our method on the simulated SAR images and several real SAR images.We also contrast it with other methods based on the same image,and the result confirm that the method proposed in this paper can complete SAR image classificatio n via a small number samples with known tags.