Research On Target Recognition In SAR Image Via Manifold Learning

Synthetic aperture radar(SAR)is a kind of high resolution imaging radar with active working model.It could work in 24-hour a day and all-weather,and provide high resolution images.With the development and maturity of SAR imaging technology,the resolution of SAR images is higher and the scene information is more detailed,which has been widely used in mineral resources exploration,disaster detection and prevention,global climate indicator monitoring,and target classification and identification in military and civilian fields.Target recognition technology is an important issues in SAR image interpretation.Manifold learing,as a data analysis tool,can find the intrinsic structural properties of images from the high-dimensional space to low-dimensional space.It could effectively reduce redundant information of images and improve computational efficiency.Therefore,this dissertation based on the manifold learning theory,this dissertation conducts research on SAR image target recognition technology,aiming to improve the accuracy and robustness of target recognition.The content and contribution of this dissertation are as follows:1.Proposing a papametric supervised manifold learning method for SAR target recognition.The traditional manifold learning method,t-stochastic neighbor embedding,(t-SNE)could not describe the siminarity between SAR images accurately.To overcome the problem,aspect angles of targets are modeled to supervise and construct the distibutional properties of the traning data in the original space.Then,an explicit nolinear mapping using kernel tric is proposed by an extension of non-parametric t-SNE for solving the “out-of-sample” problem.So the intrinsic local structure of the targets of SAR images can be obtained.By analyzing the feasibility and reasonability of the sparse representation theory for measuring the distance of manifolds,the sparse represention classification is used to calssify the low-dimensional features.Compared with the traditional manifold learning methods,the proposed method can improve the accuracy of target recognition effectively.2.Proposing a robust locality discriminant projection(RLDP)method for SAR target recognition.Although locality preserving projection(LPP)has achieve an impressive performance on SAR target recognition,the unsupervised optimization and being sestive on neighbor selection may lead to undesired recognition results.the supervised locality preserving projection is introduced to learn a linear projection,with which the SAR image can be cast into an implicit feature space.Then,we extend t-SNE to a parametric framework for optimising the linear projection.The discriminative ability of features is enhanced.Therefore,the recognition performance is improved.In addition,the proposed method RLDP is robust to the change of the neighbour parameter and does not rely on any preprocessing procedure.Moreover,to handle the non-linear geometric structure of samples,we develop a useful variant of RLDP named kernel RLDP(KRLDP)to preserve the locally non-linear discriminant structure characteristics.Thus the reliability of target recognition results is further improved.3.Under the assumption that the SAR image are lying on or close to multiple manifolds,two feature representation methods based on low-rank approximation are proposed.The JMLMA method describes the multi-manifold structure of SAR image by combining two traditional manifold learning methods;The MLA method constructs the multi-manifold structure with the pairwise similarity and local linearity rules.The low-dimensional representation model is proposed via incorporating multi-manifold regularization term into the low-rank matrix factorization framework.By alternately optimizing the matrix factorization and manifold selection,the feature representation model can not only acquire the optimal low-rank approximation of original samples,but also capture the intrinsic manifold structure information.Then,to take full advantage of the local structure property of features and further improve the discriminative ability,local sparse representation is proposed for classification.Extensive experiments demonstrate that the effectiveness of the proposed method for dealing with the target recognition under extended operating conditions.4.Considering the target recognition under the conditons that the testing samples has insufficient respresentative,a novel method named sparse and dense hybrid representation of monogenic signal is proposed for SAR target recognition.Different from the previous works only using the intensity of SAR images,the monogenic signal is adopted for obtaining the information of SAR image(energy information,structure information and geometric information)by using Log-Gabor filter bank.The formed monogenic signal feature matrix is decomposed into clas-specific dictionary and a non-class-specfic dictionary.The query sample is represented by the non-class-specific component of other class and their class memberships do not participate in the classification competition,which ensures the reliability and stability of the results of target recognition.And the experiments under extended operating conditions including configuration variation and multiple version variation verify the validity of the proposed method.