Research On Automatic Target Recognition Of Synthetic Aperture Radar Images Under Extened Operating Conditions

Synthetic aperture radar(SAR)automatic target recognition(ATR)is one of the key techniques in SAR image interpretation.However,due to the complexity of real battlefield,many measured SAR images are different with the database of the targets' charateristics.As a result,the extended operating conditions(EOCs)have been the important but difficult issues in SAR ATR.The present SAR ATR methods do not make comprehensive and systematic analysis on EOCs so they can hardly work well under these conditions.Based on these considerations,this dissertation focuses on the SAR ATR problems under EOCs in both template-based and model-based way.In the template-based way,the attributed scattering centers are adopted as the local describtors to sense the local variations in SAR images caused by EOCs.In addition,condisting the uncertainties in the test sample,the global and local features are combined to fuse their advanategs as for SAR ATR in order to improve the performance.In the model-based way,the 3-D scattering center model buit by the forward method is employed to describe the targets' characteristics.On one hand,the model is used to partially reconstruct the target owing to the flexibility of the model.And the test sample is related to the model via the similarity evaluation between itself and the reconstructed image.On the other hand,the physically relevant attributes contained in the model are introduced into the similarity evaluation between the test sample and model.Based on different types of similaries,the target type can be determined and the target's states can be referred to some extent.The main works and the innovations of the dissertation are listed as follows:1.A SAR target recognition method based on matching of attributed scattering cneters is proposed.Firstly,a robust distance measure for individual attributed scattering centers is designed,which comprehesively considers the extraction errors and the relative amplitudes of different scattering ceters.Afterwards,the Hungraian algorithm is adopted to build one-to-one correspondences between two attributed scattering center sets,where both the false and missing alarms are considered.Based on the correspondences,a similarity measure is designed,which make comprehensive anlaysis on the matched scattering center pairs,false and missing alarms,and the structure information contained in the scattering center set.Finally,the target type is determined based on the defined similarity measure.The MSTAR dataset is used to test the proposed method under different operating conditions.The experimental results show that the proposed method could achieve good performance under different operating conditions and keep robust under EOCs.2.A SAR target recognition method via joint matching of global and local features is proposed.The global and local features have different advantages for SAR target recognition.In order to combine their advantages,the original SAR image and its attributed scattering centers are taken as the global and local “filters”,repectively.The image correlation is used as the global response.Based on the one-to-one correspondence,the attribute differences between the matched scattering centers are used to calculate the local responses.Then,the random weight matrix is used to fuse the global and local responses via the linear weighting.Based on the statistical analysis on the fused results,a judgement variable is defiend for target recognition.Based on the experimental results on the MSTAR dataset,the proposed method maintains good performance under different operating conditions and work robustly under EOCs.3.A SAR target recognition method is proposed via hierarchical decision fusion of global and local features.In the practical applications,there is littlt prior information abour the operating conditions of the test sample.In order to combine the effectiveness and efficiency in a SAR target recognition method,the hierarchical decision fusion is designed to inherit the high efficiency of global features and robustness of local features to different EOCs.For the test sample,it is first fastly classified by Sparse representation-based classification(SRC)based on the random projection features.Based on the output reconstruction errors,the reliability of the present decision is calculayed.If the decision is judged to be reliable,the target type id determined.Otherwiese,the test sample is further classified based on the matching of attributed scattering centers.Based on the experimental results on the MSTAR dataset,the proposed method maintains good performance under different operating conditions and work robustly under EOCs.4.A SAR target recognition method via target reconstruction based on 3-D scattering center model is proposed.3-D scattering center model is capable of predicting partial characteristics of the target using a part of the scattering centers.This method first uses the efficient neighbor matching to build the correspondence between the attributed scattering centers from the test sample and the model predicted 2-D scattering centers.Afterwards,the 3-D scattering center model is used to reconstructe the target using the matched model scattering centers.Finally,the similarity between the test sample and 3-D scattering center model is evaluated as the image correlation between the test sample and the reconstructed image.The electromagnetic simulated data of three MSTAR targets is used to evaluate the proposed method and the results demonstrate its validity and robustness to EOCs.5.A SAR target recognition method is proposed by part-level reasoning based on 3-D scattering center model.In the 3-D scattering center model built by the forward method,each scattering center has clear physical meanings and closely related to the target's structure.In the proposed method,the point matching of scattering centers is first converted to the region matching problem and similairy between the test sample and 3-D scattering center model is evaluated based on the results.During the similarity evaluation,the spatial constraint and physical attribute constraint are introduced to obtain more robust similarity measure.Finally,the target type is determined based on the similarity measure and the target's states can also be referred to some extent.The electromagnetic simulated data of three MSTAR targets is used to evaluate the proposed method and the results demonstrate its validity and robustness to EOCs.