Research On Target Recognition And Related Processing Techniques For High Resolution Synthetic Aperture Radar Of Ground Targets

Synthetic Aperture Radar(SAR)has advantages of performing remote sensing tasks all-time and all-weather,an increasing number of radar Automatic Target Recognition(ATR)researches are conducted based on SAR images.Typical SAR ATR method for ground targets can identify interested ground tactical targets,(e.g.tanks,howitzer,armored car,and etc.),which is an effective method for the identification of friend or foe,and an important prerequisite for conducting the precise attack.SAR ATR has became increasingly popular due to the increase of SAR image resolution.Resourceful information can be provided by the high resolution SAR images,however new problems are also raised.Specifically,high resolution images show different statistic characteristics,more detailed features,greater amount of data in comparison with the medium or low resolution SAR images.Moreover,target appearances of SAR images are greatly affected by the relative pose between the SAR and the target,the radiation characteristic,geometrical characteristic,clutter,and etc.,which increased the difficulty of the recognition of SAR images.Therefore,it is important to design effective pre-processing,feature processing and classifier training methods based on the characteristics and information of high resolution SAR images.This paper aims at designing a SAR ATR method with high recognition accuracy,where several methods are proposed to improve the performance of the stages of pre-processing,feature extraction and classifier training.The performance of the proposed methods is verified with the high resolution images of the Moving and Stationary Target Acquisition and Recognition(MSTAR)public release dataset and the OKTAL simulation dataset.The main work of this paper is summarized as follows.1.For the clutter suppression task in SAR images,a clutter suppression method called Shedding Irrelevant Patterns(SIP)is proposed based on the Appearance Conversion Machine(ACM).SIP is conducted based on the analysation of the spatial distribution pattern(energy intensity and distribution of image pixels)in SAR images,where several regression functions are trained for suppressing the clutter while protecting the interested target unchanged.Then,the target image is used as the input of the trained regression function to achieve the clutter suppressed image.Experimental results with both real and simulation data verify the effectiveness of the proposed SIP method in clutter suppression tasks and the effectiveness of involving SIP in improving SAR ATR accuracy.2.For the image segmentation task in SAR images,a image segmentation method called Promoting Irrelevant Patterns(PIP)is propo sed based on the ACM.PIP analyses the spatial pattern in SAR images and calculates its similarity with the reference clutter images,a mask image is then created based on the similarity for segmentation.Experimental results verify the effectiveness of the PIP method.3.For the pose estimation and rectification problem in SAR images,a pose estimation method is proposed based on the geometrical information in SAR images.The proposed method finds the minimum bounding rectangle around the interest target and uses the estimation method that can achieve higher accuracy according to the completeness of the boundary shape of the target,e.g.the minimum bounding rectangle based method or the Radon transform based method,such that better estimation result can be achieved.Finally,the pose rectified SAR images can be achieved by rotating the original SAR image is rotated according to the estimated poses.4.For the feature extraction problem in SAR ATR tasks,a feature extraction and processing method is proposed such that a shift invariant feature set with several different discrimination features can be achieved.The proposed method extracts and combines several features(horizontal edge feature,vertical edge feature,and down-sampled texture feature)with the 2-D wavelet decomposition technique,such that a rich feature set with characteristic of shift invariance can be constructed.Then a compact feature set is constructed by using the redundancy reducing technique.In practice,different kind of wavelet basises extract various features with different classification ability.Therefore,the wavelet basis that achieves the maximum average between-class distance and minimum average inner-class distance is selected based on the Maximum Margin Criterion(MMC),such that a proper wavelet basis can be used for SAR ATR tasks,where a small average inner-class distance represents the recognizability of the same class targets and a large average between-class distance represents the high difference among the targets of d ifferent classes.5.The SAR ATR classifier is hard to train due to the high variation of target appearances and the small number of training images.A complete SAR ATR scheme is proposed,which combines various algorithms(e.g.pre-processing algorithms,feature extraction and processing algorithms,and classifier training algorithms)for improving the effectiveness of image understanding,achieving features designed for SAR images,and constructing classifiers for small sample problems,such that higher recognition accuracy can be achieved.In terms of the classifiers,several discrimination tree base classifiers are trained with both positive and negative samples,then a strong classifier is constructed by combing these base classifiers with the Real-Ada Boost framework.Experimental results with MSTAR real dataset and OKTAL simulation dataset verify the advantages and effectiveness of the proposed scheme.