Multiscale Kernel Dictionary Learning Combined with Labels for SAR Image Classification

In this paper, a novel supervised dictionary learning method, called multiscale kernel dictionary learning, is proposed for synthetic aperture radar (SAR) image classification. Due to the speckle noise in SAR images and limited training set size, we use Gaussian function with different parameters to extract an SAR image's multiscale features and exploit the principle component analysis (PCA) to reduce the dimension of these scale features. The nonlinear dictionaries are learned by introducing the dimension-reduced features into kernel dictionary learning via a nonlinear mapping function. In addition, a classification model is learned simultaneously. Hence, the objective function contains the nonlinear reconstruction error terms and a classification error term. An incremental method and the kernel orthogonal matching pursuit (KOMP) algorithm are used to solve the optimization problem. Experimental results on the MSTAR dataset demonstrate that the proposed algorithm outperforms some representative dictionary learning methods, especially under small training set size condition.