SAR Images Target Recognition And Change Detection Based On Sparse Feature Learning

Synthetic aperture radar(SAR),which can acquire the high resolution images,is not be limited by the light and weather conditions.Therefore,it gains widespread application in various civilian and military fields.Automatic target recognition(ATR)and change detection are two significant components in SAR image understanding and interpretation.The researchers and scholars all over the world are concerning about these.On account of its coherent imaging system,SAR images have the speckle noise inevitably,which has become one of the problems in SAR images processing.In order to suppress the speckle noise,the thesis takes the advantages of the spatial pyramid matching(SPM)and sparse representation to learn the discriminant features in images.And then these features be used to research and analysis the ATR and change detection.There are five main works of this thesis which is summarized as follows:1.The presence of the speckle noise may hinder the targets in the SAR images.A novel sparsity weighted method is proposed for SAR image ATR.First,an image is divided into gradually fine sub-regions.Next,the feature vector of the sub-regions at every pyramid level is weighted on the basis of the dependability which is determined according to the residuals obtained by sparse representation.Finally,the representation for SAR images will be built through systematically concatenating the weighted feature vectors.With sparse representation classification,this method could enhance the weights to the pooling features generated in sub-regions located in the target and suppress the weights of the background.The experimental results verify the superior performance of the proposed method in dealing with the speckle noise and large clutter background.2.A Complementary Spatial Pyramid Coding(CSPC)approach is proposed.Some discriminative information is easy to loss in the SPM coding process,which limits the representation capability.In order to improve the discriminative power of the learned local features,the coding residual is encoded for capturing the lost discriminative information and plays a complementary role to the original local feature.The recognition power of combination of them would outperform each individual of them.Because the synthetic aperture radar(SAR)images are sensitive to the target aspect angles,the variation of the intra-class may be larger than the inter-class.To reduce the interference of noise and get more discriminative local image representation,we construct a codebook for each class which utilizes the prior label information in the step of sparse coding.The experimental results verify the concatenation of these two kinds of features are improved effectively.This well verifies that the residual feature representation is discriminative,salient and complementary to the local feature of SAR image representation.The recognition performance based on the proposed method is compared with some other feature extraction methods and the proposed method obtains a better accuracy.3.The available feature extraction techniques for change detection always ignore the spatial context correlation and are not robust to speckle noise.To overcome these drawbacks,we present a novel feature extract technique which takes full advantage of sparse representation and non-local similarity in the framework of SPM.This sparse coding technology could suppress the speckle noise effectively.Taking the advantages of the non-local similarity,we search for the most similar pixels on the whole image plane to generate a group of feature vectors for each pixel.The pooling method is applied to obtain the feature vectors groups,and then get the final discriminative features.The pooling enhances the similarity of homogeneous areas and reduces the correlation of the heterogeneity in a latent way.The experimental results verify that the proposed method can identify the changed areas and the unchanged areas on several real SAR image data sets and simulated image pairs.4.An unsupervised feature learning method,which applies the non-negative matrix factorization(NMF)algorithm and the non-negative sparse coding algorithm,is presented for SAR images change detection.The non-negative matrix factorization(NMF)method is first used to obtain the dictionary which could contain spatial structure information.Then,we extract the feature for each pixel apply the sparse coding to increase the discrimination of the feature.Finally,the result of SAR image change detection is generated by applying simple k-means clustering which could divide the learned features into two clusters.Because the data in the SAR images are non-negative,non-negative matrix factorization(NMF)guarantees this property and the basis matrix obtained from NMF contains the local construction information of the difference image.Experimental results confirm the effectiveness of the proposed method on real SAR image datasets which have different levels of noise.5.A novel change detection approach for SAR image is proposed to improve the performance of the unsupervised methods.It is based on the saliency similarity detection and the semisupervised Laplacian support vector machine(SVM)to classify the changed areas and the unchanged areas.A pseudotraining set is generated according to the saliency similarity detection.The saliency similarity could obtain the correct changed and unchanged areas which is proved in the experiment.The Laplacian SVM explores the prior information of the available labeled samples and combines unlabeled samples to enhance its discrimination.Experimental results on several real SAR image datasets confirm the proposed method could increase the accuracy detection rate for change detection.