Study On Feature Extraction And Recognition Of Air Targets Using ISAR Images

Inverse Synthetic Aperture Radar (ISAR) can obtain two-dimensional highresolution images, with detailed information, such as scale, shape, structure and gesture,which afford abundant features for classification and recognition of targets, in allweather, day or night. Therefore, the technique of automatic target recognition (ATR)using ISAR images is playing a very important role in modern military applications andis always being paid attention to. In order to resolve the problems of classifying airtargets using ISAR images, several special researches are made on the basis ofgeometric modeling and imaging simulation in this dissertation, including ISAR imagespre-processing and feature extraction algorithms of ISAR images.In chapter1, the research background and significance is introduced, and thedevelopment of ATR techniques using ISAR images is reviewed. Then aerodynamiclayout of air targets, scattering properties of aircraft components and the key in ATR ofISAR targets are analyzed in detail, followed by the introduction of main content in thisdissertation.In chapter2, modeling and simulation of ISAR targets, image pre-processingtechnology are inverstigated. At first, the principles of ISAR imaging, Open Flightmodeling with Mu1tigen Creator, simulation of ISAR imaging using Matlab with radarwave generated by RadBase are introduced; then, the inherent characteristics and mainfactors affecting ISAR images are analyzed, followed by correlation stduy betweenISAR images using F14and F18aircrafts as examples; finally, suppression of thespeckle noise and interferential stripes，images normalization are proposed to enhancethe stability of images，which is propitious to the latter feature extraction andrecognition.In chapter3, algorithms for aspect estimation and affine invariant featureextraction are presented. First of all, in order to improve the efficiency of templatematching, we combine the thoughts of target’s principal axis extraction and minimumenclosed rectangle acquirement, and propose a correlation matching method based onaspect estimation, which narrows the scope of template search by aspect estimation andimprove the efficiency of ISAR template matching; secondly, aimed at resolving theproblems of translation, gesture and scale inconsistencies, we extract key points ofISAR images, constructs affine-invariance features which are then saved in the hashtables, and uses geometric hash as the final classifier. Simulations show the presentedalgorithm not only effectively distinguishes targets with different structures, imagingperspective and scale, but also performs excellently in local recognition.In chapter4, algorithms for statistical feature extraction of ISAR images based onnon-negative matrix factorization (NMF) are proposed. Introducing visual perceptionconcept to classification, we adopt NMF as a mathematical tool to obtain the statistical feature vector of an ISAR image, and propose approaches of feature extraction based onoptimized discriminant non-negative matrix factorization and subclass discriminantnon-negative matrix factorization. At first, we construct a feature space based on localnon-negative matrix factorization (LNMF) and non-negative sparse coding (NNSC),which is convenient for local and sparse feature extraction, respectively, then optimizethe feature space by screening the feature base to enhance the classification performance,and finally present algorithms for ISAR target recognition based on optimized discri-minant local non-negative matrix factorization (ODLNMF) and optimized discriminantnon-negative sparse coding (ODNNSC); secondly, to deal with the multimodal distribu-tions of ISAR images, constraints inspired by the clustering based discriminant analysis(CDA) are imposed on LNMF and NNSC to deduce algorithms for ISAR target recog-nition, which are based on subclass discriminant local non-negative matrix factorization(SDLNMF) and subclass discriminant non-negative sparse coding (SDNNSC).SDLNMF and SDNNSC extend the LNMF and NNSC decomposition, respectively, byembedding the subclass discriminant constraints, and reformulate the cost function toachieve discriminant projections. Simulations demonstrate the effectiveness and ro-bustness of the proposed method.In chapter5, algorithms for manifold feature extraction of ISAR images based onorthogonal neighborhood subclass discriminant projections (ONSDP) are presented. Atfirst, ONSDP is proposed to extract low-dimensional feature from high-dimensional andmultimodal ISAR images, which incorporates local topology preserving and subclassdiscriminant dimensionality reduction. Moreover, an orthogonal constraint is introducedto preserve geometric structures of ISAR images; secondly, to explore nonlinear struc-ture of ISAR targets, we add kernel function to ONSDP and propose kernel orthogonalneighborhood subclass discriminant projections (KONSDP), which replaces innerproduct with RBF kernel function and maps implicitly original nonlinear data to a high-er dimensional linearly separable space. Derivation process and verification by simula-tions are given later; thirdly, regarding an ISAR image as the second-order tensor, wepropose algorithm steps as well as derivation process of tensor subclass discriminantanalysis (TSDA) and tensor orthogonal neighborhood subclass discriminant projections(TONSDP), which preserve the structure information between the rows and columns inthe ISAR image and takes advantage of subclass discriminant analysis. Optimization oftensor subspace is finally converted to the problem of generalized eigenvectors Simula-tions verificate the advantage of these methods in target recognition rate.Charpter6summerizes this dissertation and discusses the future work.