Novel hybrid methods based on pca for radar target recognition and classification

In radar imagining of the targets, Synthetic Aperture Radar (SAR) and Inverse Synthetic Aperure Radar (ISAR) systems are used. For recognition and classification of the targets, many methods exist in literature and many of them require Radar Cross Section (RCS) data of the targets. RCS data are obtained by using electromagnetic wave. Incoming electromagnetic wave reflects from target surface and the scattering electromagnetic wave generates eletricity field. For obtaining of the RCS data, scattering electromagnetic wave is calculated. Angular diversity and frequency diversity techniques are used to enhance RCS data. In angular diversity, frequency is constant, aspect and azimuth angles are changed in an interval. In frequency diversity aspect angles are constant, the frequency is changed in the interval. In this way, RCS data of targets are formed as a matrix. For obtaining RCS data of targets, RCS formulas of the elliptical cylinder, rectangular flat and isosceles triangular flat are used. Two main data set are obtained for angular and frequency diversity. Afetrwards, using same and different targets, other two main data set are obtained. Totally four data set are generated for both recognition and classification of radar targets. Supervised and unsupervised methods are main investigation topics in classification and recognition. In supervised methods, classification and regression are achieved with using known class label. On the other hand, in unsupervised methods clustering of the data is obtained without using class label. In training stage of this work, known class labels are used and feature vectors are generated. Thus recognition of classes is possible with supervised methods. In testing stage of this work, class labels are unknown and classification is achieved via unsupervised methods. Similarity between test and training feature vectors of targets is calculated for decision of class. Similarity means euclidean distance between feature vectors of targets. Therefore minimum distance to training feature vectors determine class of the unknown test targets. In this work, feature extraction based radar target recognition methods are investigated. Main difficulty of the methods is selecting features in terms of best classification performance. In literature there are different aproaches for choosing features. For classification of radar targets, Principal Component Analysis (PCA) based approaches are examined. Kernel PCA, Linear Discriminant Analysis (LDA) and Fisher Linear Analysis methods are also investigated in this work. By using RCS matrix, performance comparison of the methods are compared with each other. Moreover, hybrid form of the PCA, KPCA, LDA and FLD are proposed. These forms are PCA+LDA, PCA+FLD, PCA+LDA+FLD, LDA+FLD, KPCA+LDA, KPCA+FDA and KPCA+LDA+FLD. PCA projects multi dimensional data to less dimensional space. To perfom projection, linear and orthogonal transition is required. The covariance matrix helps linear and orthogonal transition in PCA. Principal components are obtained through the eigenvalues and eigenvectors of covariance matrix. By the time, data compression is achieved via PCA. PCA based methods stand on chosing largest eigevalues and eigenvectors. However chosing process of eigenvalues and eigenvectors is not based on optimum criterion. In PCA method, data has to be distributed with high variance to classify well. In distributed data with low variance, PCA is inefficient and classification performance decreases. In this sitution PCA can not seperate classes on subspace by using linear transition. Therefore KPCA methods have become important in terms of non linear transition. KPCA projects data to subspace with using nonlinear transition. To perform nonlinear transition, kernel fuctions are used. In literature, gauss, polynominal and hyperbolic kernel fuctions are mostly used. In this work, polynominal kernel fuction is investigated and used for recognition and classification of radar targets. In chosing features vectors, FLD method brings different approaches to best seperation using Fisher criterion. Many developped method exist for solving Fisher criterion. In litearture Classical Linear Discriminant Vectors (CLDV), Classical Linear Discriminant Vectors with Whitening Process (CLDVM), Weighted Pairwise Fisher Criteria Vectors with Whitening Process (WPFCVW) techniques are studied. In this work, CLDV technique is considered. Thus optimum eigenvalues and eigenvectors are chosen. PCA+LDA is one of the frequently used hybrid method in literature. Because PCA has disadvantages for seperation of same classes with low variance, PCA+LDA takes advantage by increasing successful recognition rate. Meanwhile, novel methods: PCA+LDA+FLD, LDA+FLD, KPCA+LDA, KPCA+FLD and KPCA+LDA+FLD are proposed in this work. This new methods have significant successful recognition rate for specific situations. MATLAB codes are also developed to compare the methods with each other. Moreover, Matlab Graphical User Interface (GUI) program is designed. Therefore setting of eigenvector number, signal to noise ratio (SNR), size of data sets and noise power make the performance analysis easy. Predicted success percentage is shown in a table. As a conclusion, novel methods are compared successfully. Successful recognition rates of methods are calculated and used for comparisons. PCA has highest successful recognition rate on classes that has high variance. FLD has highest successful recognition rate with %100 for frequency diversity. When examined PCA+LDA+FLDA and PCA+LDA methods for angular diversity and target of different class, successful recognition rate has increased for lower eigenvalues used. FLD method has incresed successful recognition rate according to high eigenvalues used. PCA+FLD method has increased successful recognition rate for big size of training data vectors. KPCA+LDA has the lowest successful recognition rate with %44. LDA+FLD has the lowest successful recognition rate with %33,33. For the targets of same class , KPCA, KPCA+LDA, KPCA+FLD and KPCA+LDA+FLD methods have the worst successful recognition rate comparing with the PCA, PCA+LDA, PCA+FLD, PCA+LDA+FLD, FLD and LDA+FLD methods for frequency diversity technique. The result also show that under the certain condition methods performs desired recognition rate.