Study On The Methods Of Feature Extraction And Recognition Of Ships In SAR Imagery

Independent with weather and illumination conditions, synthetic aperture radar(SAR) can observe the ocean area with the characteristic of macro-scale, long term,continuous and real-time. Along with the development of SAR imaging and SAR imageinterpretation technology, ship recognition in SAR image attracts much attention andbecomes an important research aspect of ocean surveillance. this dissertationinvestigates the main procedure of ship recognition in SAR imagery, including targetsegmentation, feature extraction, feature selection and the designing of classifier.One of the basic procedure for ship recognition is to segment out the target area inSAR imagery accurately. Aiming at the phenomenon of azimuth ambiguity and rangespreading, a refined ship segmentation method based Radon transform is proposed. Themethod separates the ship from its outspread shadow or cross shadow in radontransformed domain firstly, and then segment the ship itself in SAR image domainutilizing2-D Otsu algorithm. Evaluation results on measured SAR imagery reveal thatour method outperforms the Otsu algorithm.Another important phase of ship recognition is feature extraction. The dissertationextracts the geometric structure and intensity statistical features firstly. And thenanalyses the ship’s macro-scale geometric structure and its electromagnetic scatteringmechanism in depth. Finally, a novel electromagnetic scattering feature named localRCS density is proposed. The proposed feature reflects the relationship between ship’smacro-scale geometric structure and the corresponding scattering energy distribution inSAR imagery. The novel feature provides better discrimination ability of different kindsof ships by taking into account the local structure information.Feature selection and the design of classifier are also very crucial to shiprecognition in SAR imagery. In one hand, we analyses the filter-based andwrapper-based feature evaluation strategy and proposes a serial feature selectionalgorithm for the problem of feature selection. The algorithm prescreens the featurewith multiple rules of filter-based evaluation strategy and refines out an optimal featureset with wrapper-based strategy. While integrating the advantages of the filter-based andwrapper-based strategy, the algorithm reduces both the dimension of the optimal featureset and the consumed time. In the other hand, the dissertation designs a support vectormachine (SVM) combination classifier on the basis of analyzing the performance of therepresentative classifiers such as K-nearest neighbor (KNN), Bayes, back propagation(BP) neural network and SVM, et al. The proposed classifier make full use of differentinformation of three individual classifiers, including KNN, Bayes and BP neuralnetwork by a nonlinear combination based on SVM. Experiment results conducted onmeasured TerraSAR-X SAR imagery validates the effectiveness of the proposed feature selection algorithm and the combination classifier based on SVM.