Spatial Mismatch Correction And Classification Of SAR Target Based On Three-Dimensional Electromagnetic Scattering Model

Synthetic Aperture Radar(SAR)has the advantages of all-day,all-weather and penetrability,which has made it an important method of enemy target surveillance and battlefield reconnaissance.Traditional SAR image interpretation methods rely on manual interpretation and cannot meet the real-time requirements of the battlefield.According to SAR target identification data sources,the current identification methods can be divided into template-based methods and model-based methods.The template-based method is constrained by the limited number of the collected template images.Based on the model,the model-based approach can obtain omni-directional SAR images and can flexibly simulate target structure changes,which has more advantages than the template-based approach.In the model-based method,the three-dimensional electromagnetic scattering model has both advantages of the model and rich electromagnetic properties,which is the focus of this paper.The research in this paper mainly focuses on the three-dimensional electromagnetic scattering model,and the specific research work is as follows:Firstly,the research significance,status and relevant knowledge of three dimensional electromagnetic scattering model for SAR target recognition are summarized.The development of the model-based target recognition method and the problems in SAR target recognition are introduced.Secondly,based on the space mismatch problem,the correction method is proposed.Due to the difference between the SAR system that constructs the three-dimensional electromagnetic scattering model and the SAR system that collects the target data to be measured,there is error between the spatial coordinates of the target to be measured and that of the model.Therefore,there is a spatial mismatch problem between the 3d electromagnetic scattering model image and the target image to be measured.To solve this problem,a space mismatch correction algorithm is designed in this paper.Based on the scattering center feature,the proposed method utilizes Hausdorff distance to measure the mismatch degree between the electromagnetic scattering model image and the measured image,and by searching for the optimal parameters to minimize the mismatch,the optimal spatial transformation of the corrected model image is obtained.The mismatch simulation environment is built from the measured data in the MSTAR data set.Experimental results show that the proposed algorithm is feasible and effective under standard and extended operating conditions.Thirdly,three-dimensional electromagnetic scattering model-based region matching and target recognition method is designed.In the proposed method,the three dimensional scattering center in the model is projected into a binary region,and then matches with the test sample's target region.In order to build the association between model and test sample for target recognition The similarity criterion is defined based on the result of region matching.In the experiment,the electromagnetic simulation data of three types of MSTAR targets are utilized to test the performance of the proposed method and compared with the two classic template-based methods.Experimental results show that the proposed method can maintain a well recognition performance under standard operating conditions and various extended operating conditions.Finally,a deep learning network based SAR target image recognition method is constructed.Based on the deep residual network,the proposed method learns invariant features of target characteristics from MSTAR data set,and In the feature space extracted from the deep learning network model,all classes' center are established.When to identify the test image sample,based on the deep learning network model,the test sample is transformed into the space where the class centers in,and the similarity between the test sample and class center is calculated.The proposed method has no requirement for the professional knowledge to describe the SAR target characteristic,but only needs to utilize the data-driven approach to obtain better identification results,which reduces the time consumption for modeling the recognition target.The proposed method has achieved 99.65% classification accuracy on three target data sets of MSTAR.In addition,the experimental results of the proposed method in the electromagnetic simulation data of three types MSTAR targets also reach a high recognition accuracy,which verifies the effectiveness of the method.