Study Of Feature Extraction And Target Recognition For SAR Images

As an effective means of earth observation,synthetic aperture radar(SAR)is playing an increasingly important role in military reconnaissance,topographic survey and marine monitoring.The inherent characteristics of the SAR image and the expansion of the data scale,however,lead to a great challenge on SAR target recognition by manual way.SAR automatic target recognition(ATR)has received general concern.This paper focuses on the methods of feature extraction and target recognition for SAR images: Considering the limitation of target recognition with a single kind of feature,we put forward a SAR ATR method which takes advantages of two kinds of features,including the amplitude feature and the scale invariant feature transform(SIFT)feature.Of which the amplitude feature describes intensity information and the SIFT feature describes the gradient information and they can be complementary to each other.What's more,the point-wised gated Boltzmann machine(PGBM)is introduced into SAR ATR.In order to make it more suitable for real valued data,the conditional probability distribution of the visible nodes is extended from the Bernoulli distribution to the Gauss distribution.The main contents of each part can be summarized as follows:1.According to the different recognition criteria,the existing SAR image target recognition methods are divided into three categories,which are the methods with the matching criteria,the reconstruction criteria and the classifier.Among the existing methods,the template matching method,the point set matching method based on HD distance,the target recognition method based on PCA transform and linear SVM classifier and the sparse representation method are introduced.The performance of each method in SAR image target recognition is verified by experiments,and the advantages and disadvantages of each method are analyzed.2.As the globalfeature cannot distinguish foreground and background and is easily affected by defects occlusions,we introduce the SIFT feature into SAR ATR.We first study the target recognition method of SAR image spatial pyramid model(SPM)and the bag of words(Bo W)model based on the densely extracted descriptors to validate the effectiveness of the SIFT feature applied to SAR ATR.Furthermore,according to the complementarity of gradient information and amplitude information,a novel method which takes advantage of boththe SIFT feature and the amplitude feature is proposed.Specifically,the global SIFT features and the amplitude features of training samples are first put together respectively to construct dictionaries,then the joint dynamic sparse representation(JDSR)model is introduced to combine the two features and at last the recognition results are obtained according to the reconstruction error.Experimental results on the MSTAR dataset validate the effectiveness of the proposed method.3.We propose a new SAR ATR method based on the Gauss Point-wise Gated Boltzmann Machine(Gaussian PGBM).On the basis of readily introduction of the restricted Boltzmann Machine(RBM),the Point-wise Gated Boltzmann Machine(PGBM)is introduced to SAR ATR.Compared to the RBM,the PGBM introduce a switch unit for every visible node to unify the nodes selection and feature extraction into one model.In order to make it more suitable for the modeling of real value data,the conditional distribution of the visible nodes is extended from the Bernoulli distribution to the Gauss distribution.Experimental results on the MSTAR dataset show that the model can effectively select the target pixels from the background,which has a certain application value.