Study On SAR Image Target Detection And Recognition Under Complex Background

The Synthetic Aperture Radar(SAR)is an irreplaceable instrument in information acquisition with the advantage of working in day/night,all-weather,high resolution and long range exploration.Therefore,SAR plays an important role in many military and civilian fields.With the rapid development of SAR imaging technology,the capability of image collection has grown much stronger.How to distinguish target from the sea-like background data and classify the target automatically is an urgent mission.The work of this dissertation focus on two aspects,the first one is SAR target detection under complex background,and the second one is the SAR target recognition.For SAR target detection,the underlying assumption of the RCS of target is larger than clutter may not be satisfied,especially in the weak target case.In this dissertation,instead of detecting based on energy,polarimetric information based methods are proposed.For SAR automatic target recognition,to deal with the problem of high-dimensional,multi-modal distribution resulting from the azimuth sensitivity,supervised dimensional reduction methods are proposed.The main contents of this dissertation are summarized as follows:1.Aiming at the problem of energy dependence in traditional target detection,a novel algorithm of supervised incoherent dictionary learning for ship detection with Pol SARimages(SIDL)is proposed.Considering that the polarimetric scattering mechanism difference between man-made target and nature clutter,in the dictionary learning stage,SIDL increases the discriminancebetween target and clutter via decreasing the sub-dictionary reconstruction ability to the cross-samples and penalizing the sub-dictionary cross-incoherent.A high detect probability performance is achieved with a low false probability.The algorithm make use of only the polarimetric information,therefore,the weak target can be well detected.The performance is validated with the RADARSAT-2 dataset.2.Considering that the scattering mechanism of target is various,an algorithm of Pol SAR ship detection based on multiplepolarimetric scattering mechanisms(DPMLVSVM)is proposed.Resulting from the multiple scattering mechanisms,the samples distribute according to a multi-modal distribution,therefore,a poor detect performance would be obtained via a single scattering mechanism based method.Taking the advantage of non-parametric,DPMLVSVM split the data space into a set of local regions and learn a local polarimetric detector in each region.Assembling each local detector,a multiplepolarimetric detector is obtained.Similar to SIDL,DPMLVSVM is a polarimetic detector independent to the energy.To decrease the redundancy of the polarimetric decomposition features,a sparsity penalty is added into DPMLVSVM and a sparsity promoting DPMLVSVM(SPDPMLVSVM)algorithm is proposed.A RADARSAT-2 set verified the performance of the proposed algorithm in weak target detecting.3.Considering that small sample size and high model complexity likely resulting to overfitting,a robust max-margin linear discriminant prjection algorithm is propoesed.This algorithm learns the projection and classifier jointly under a fully Bayesian framework.A disriminant projection space is learned.With the help of noise corruption,the risk of overfitting is reduced.The performance of the proposed model is varified via the MSTAR data set.4.In order to deal with the multimodal distribution resulting from the azimuth sensitivity,supervised dimensional reduction algorithms combined with the nonparametric are proposed.In this section,two algorithms are proposed,i.e.the infinite Bayesian max-margin linear discriminant projection(i MMLDP)and the infinite kernel max-margin discriminant projection(i KMMDP).In i MMLDP,by assembling a set of local regions,where we make use of Bayesian nonparametric priors to handle the model selection problem,e.g.the underlying number of local regions.In each local region,i MMLDP jointly learns a discriminative subspace for feature projection and the corresponding classifier for classification.Therefore,under this framework,i MMLDP combines dimensionality reduction,clustering and supervised learning in a principled way.Moreover,to deal with more complex data,e.g.local nonlinear separable structure,we extend the linear projection to the nonlinear case based on the kernel trick and develop the i KMMDP model.Thanks to the conjugate property,the parameters in both models can be inferred efficiently via the Gibbs sampler.Finally,the models are implementd on synthesized and real-world data,including multi-modally distributed datasets and MSTARdata set,to validate their efficiency and effectiveness.