Research On The Pre-processing Methods For Vehicle Target Classification In SAR Images

In the last two decades,great improvements have been made in Synthetic ApertureRadar(SAR) technology. As an active radar system, SAR can penetrate the clouds and acquiredata dayand night. Meanwhile,as the developmentofSAR system, automatic target recognition(ATR) with SAR images has become an important application of SAR image interpretation.However, according to the characteristics of SAR imaging mechanisms, the geometric distortionand the speckle noise of SAR images affect the SAR applications. As a result, the traditionaloptical image processing methods can hardly be used in SAR ATR with a high accuracy and arobust performance. Therefore, some preprocessing procedures with SAR images are adoptedbefore the employment of ATR, which can restrain the speckle noise and improve theperformance effectively. Moreover, some features can be extracted through the preprocessingprocedures, such as the outline and the aspect of the targets, which can be used in the followingrecognition algorithms. Therefore, the preprocessing procedure plays an important role in SARimage interpretation, especially in the target classification based on target masks, where thedifference between the real SAR image and the mask is obvious. This paper mainly focused onthe aspect estimation of the target and the image pre-processing procedure on targetenhancement so as to improve the performance of targets classification with SAR images basedonthesimulatedtargetmasks.In theSAR ATR system, the estimation of target aspectis a very important procedure. Thecomputationefficiencyandrecognitionperformancewill be greatlyimprovedifthetarget aspecthas been estimated in advance. The aspect estimation methods based on the dominant boundary,the envelope box and combination of the two were studied in this paper firstly. Then, a novelmethod for target aspect estimation based on Radon transform of the leading edge is proposed.The new method is introduced to eliminate the ambiguity of horizontal and vertical aspectestimation usingthe length of the target. The experimental results with the MASTAR data showthevalidityandrobustnessofthealgorithm.In SAR image enhancement, a regularization algorithm based on l knorm is also validatedusing the MSTAR data,whichcan effectivelyreduce the difference between the real SAR imageand the simulated mask. The algorithm can also improve the signal to noise ratio (SNR) andestablish a well foundation for the following classification of SAR images based on simulationtemplates.To validate the effect of the preprocessing procedures in SAR target classification, twomethods were proposed based on simulation template in pixel-level and feature-levelrespectively. A contrastive experiment used the simulation data and the MSTAR data was designed, and the results indicated that the performance has been improved through thepreprocessing procedure. The algorithm proposed in the thesis is appropriate for SAR targetclassification in a real SAR application system, which can reduced the memoryof templates andthecostofdatasamplingobviously.