Target Recognition Of Coherent Ladar Range Image For Small Samples

With the development of laser technology, laser imaging radar has graduallyobtained widespread applications, and target identification technology of laser hasbecome a hot topic at home and abroad. In this thesis, our aim is to establisheffective algorithms of three-dimensional target recognition based on small samplesof laser radar (ladar) range images. The main work is divided into four parts, asfollows:In order to reduce computing amount of Zernike moment invariants (ZMIs),even-order moments of ZMIs are proposed to target recognition of ladar rangeimages in this thesis. According to target recognition of ladar range images, therelations among three kinds of moment invariants including Hu moment invariants(HMIs), ZMIs, affine moment invariants (AMIs), two kinds of classifiers includingsupport vector machine (SVM) and back propagation neural network (BPNN) arediscussed. Using the real ladar images, rotation invariance of three kinds ofmoments invariants are proved.In order to improve the performance of target recognition, four combined-moments invariants are introduced into target recognition of ladar range images. Inorder to eliminate the “Hughes” effect caused by combined-moment invariants withsmall samples, according to random subspace ensemble of SVM (RSE-SVM), semi-random subspace ensemble of SVM (sRSE-SVM) is proposed. Moreover, RSE-SVM, sRSE-SVM, Relief algorithm and SVM recursive feature elimination (SVM-RFE) are applied to select features of combined-moment invariant to improve theperformance of target recognition of ladar range images.In order to improve the performance of non-articulated target recognition withsmall samples, classifier ensembles-SVM ensembles and BPNN ensembles-areapplied to target recognition of ladar range images. Moreover, based on singlemoment invariants and combined-moment invariants, the performances amongclassifier ensembles, Relief algorithm, SVM-RFE, RSE-SVM and sRSE-SVM are compared. The generalization ability of SVM ensembles and BPNN ensemblesbased on single moment invariants are analyzed with the pitch angles of the testedsamples changing while the pitch angles of the trained samples is invariant.In order to improve the performance of articulated target recognition withsmall samples, elastic shape analysis is introduced to ladar range images based onsmall samples. Recognition rates of elastic shape analysis under arbitrary azimuthangles are analyzed and compared with three kinds of moments invariants-HMIs,AMIs, ZMIs-with SVM with different carrier to noise ratio (CNR).The invarianceof elastic shape analysis, including in-plane rotation, translation, and scaling arealso analyzed.