Study On Target Discrimination And Recognition For SAR Imagery

Synthetic Aperture Radar(SAR)is widely used in military and civil fields,owing to its all-time operation,reduced sensitivity to weather conditions,and penetration ability.With the continuous development of SAR imaging technology,airborne and spaceborne SAR system have been widely used,and obtained a large number of large scale SAR images.It is almost impossible to manually interpret these SAR images,and there is an urgent need for automatic interpretation of SAR images.A typical SAR Automatic Target Recognition(ATR)system includes three stages: detection,discrimination and recognition.There are still some problems to be solved for discrimination and recognition,so it is of great significance to study discrimination and recognition.This dissertation focuses on the discrimination of artificial clutter in discrimination stage,and feature extraction and recognition methods in recognition stage.The main content of this dissertation is summarized as follows:1.Study on the discrimination of artificial clutter in SAR target discrimination stage.The purpose of the discrimination stage is to eliminate the clutter false alarms,which are mixed in the potential target chips detected in the detection stage,and maintain the real targets as much as possible.Traditional discrimination features are difficult to effectively discriminate artificial clutter false alarm,which led to a large number of artificial clutter false alarms passed to the recognition stage,these artificial clutter false alarms may affect the performance of the recognition stage.This dissertation focuses on the problem of artificial clutter rejection in SAR target discrimination stage,and a method of SAR target discrimination based on scattering center feature and K-center one-class classification is proposed.The scattering center feature is composed of the amplitude and location of scattering centers,which makes full use of the location information of scattering center.The amplitude of natural clutter is generally different from that of target,thus the natural clutter can be discriminated from target by amplitude information.And the artificial clutter can be discriminated by the location of the scattering centers,which can reflect the topological structure of the target.The number of scattering centers extracted from different SAR chips may be different,that is,the dimension of scattering center features of different chips may be different.Thus,it is a difficult problem to measure the similarity(distance)between scattering center features with different dimensions.In this dissertation,we adopt the Hausdorff Distance(HD)to measure the similarity between scattering center features,and improve the traditional K-center one-class classification based on the HD distance.Finally,the improved K-center one-class classification is used as discriminator.Experimental results based on the measured MiniSAR dataset show that the discrimination performance of the scattering center feature is better than that of the conventional discrimination features,especially for the artificial clutter.2.Extract features related to the electromagnetic scattering characteristics of SAR targets for recognition,based on the attribute scattering center model(ASCM),from the aspect of coherent imaging mechanism of SAR images.Although the traditional recognition features have gained well performance on SAR image recognition,only limited low-level features based on the dependencies among the pixels in local region are exploited,most of these features are low-level,which is vulnerable to the influence of speckle noise.In this dissertation,a novel recognition method based on the ASCM model and discriminative dictionary learning is proposed.This method consists of three main stages.In the first stage,namely low-level local feature extraction stage,the parameters of ASCM model are selected by genetic algorithm(GA),and then the low-level local features are generated by convoluting the ASCM model with SAR chips.In the second stage,namely feature coding stage,a discriminant dictionary learning method called Label consistent and Locality constraint Discriminative Dictionary Learning(LcLcDDL)is proposed,which integrates the label information and local geometric information into the sparse codes of low-level local features.In the third stage,namely feature pooling stage,Spatial Pyramid Matching(SPM)with max pooling is used to integrate the sparse codes of low-level local features into the final high-level global features.Then,the high-level global features are sent into the linear support vector machine(LSVM)to recognize which classes the chip belongs to.Experimental results based on the Moving and Stationary Target Acquisition and Recognition(MSTAR)dataset demonstrate the separability of the sparse coefficients and the recognition performance of high-level global features.3.Focus on the problem of SAR target recognition,a new SAR target recognition feature,called LBP-Haar-like(LHL)feature,based on the Haar-like(HL)feature and the local textural information of SAR image,is proposed in this dissertation,and LHL feature is then applied to target recognition to verify its effectiveness.In SAR target classification,feature extraction plays a crucial role.HL feature is proposed based on the same principle of Haar wavelets.It is a simple and inexpensive rectangle feature,which represents the difference between the sum of pixel intensities in two or more adjacent rectangle regions.In SAR chips,the echo signal of radar waves reflected by the target vary with the change of the azimuth of the target,while the local textural information of SAR target changes little with the change of azimuth.Therefore,the local textural information of SAR target is useful for target recognition.In order to make full use of the local textural information of SAR target,and retain the simple and inexpensive characteristics of HL feature,we introduce local binary pattern(LBP)to integrate the local textural information of SAR target into HL feature,and propose LHL feature.The prototype of LHL feature is the same as HL feature,and the local textural information of SAR chip is extracted by LBP.The exhaustive set of rectangle feature of SAR chip is larger,around tens of thousands,it is not convenient for most classifiers to deal with such high dimension feature.Random Forests(RF)classifier is insensitive to feature dimension,thus,we adopt RF classifier to distinguish the classes of chips.Finally,experimental results based on the MSTAR dataset show that the recognition accuracy of LHL feature is higher than that of HL feature.