Multiple kernel learning via orthogonal neighborhood preserving projection and maximum margin criterion method for synthetic aperture radar target recognition

摘要

A multiple kernel learning (MKL) method is proposed for synthetic aperture radar (SAR) target recognition. The goal of the proposed MKL is to learn an optimal combined kernel to reduce the dimensionality of SAR images and maximize the separability of SAR targets. Orthogonal neighborhood-preserving projection (ONPP) can effectively reduce the sample dimensionality and maximally preserve the structure information but without the discrimination. On the contrary, maximum margin criterion (MMC) has the ability of classification but without the ability of preserving structure information. To realize the proposed goal, ONPP and MMC are combined within the graph embedding framework, where an optimal projective direction and basic kernel weights are automatically learned. Based on the obtained projection direction and kernel weights, all basic kernels are projected to generate the composite kernel. Moreover, the projection and transformation operations are based on three-dimensional (3-D) data generated by a series of basic kernel matrices, which can completely preserve the structure information in reproducing kernel Hilbert space. Numerical experiments on MSTAR data-set indicate that the proposed MKL method can effectively reduce the dimensionality of SAR images and achieve the outstanding recognition performance when compared with several state-of-the-art algorithms.