Synthetic Aperture Radar Target Recognition Using Weighted Multi-Task Kernel Sparse Representation

Chen Ning,Gong Zhang,Wenbo Liu,Xin Wang

doi:10.1109/access.2019.2959228

Abstract

As an extension of traditional sparse representation (SR), kernel SR has received great interest recently in the areas of computer vision and pattern recognition. It shows a considerable capacity to map linearly inseparable data into high-dimensional feature space via nonlinear mapping technique, and has been widely used in target recognition problems. In this paper, we propose a new weighted multi-task kernel sparse representation method to solve the synthetic aperture radar (SAR) target recognition problem. To capture the spatial and spectral information of a SAR target simultaneously, the proposed method explores the monogenic signal transformation to generate multi-scale monogenic features at first. Then, the proposed method provides a unified framework, named multi-task kernel sparse representation, for SAR target classification. The framework implicitly maps monogenic features into a high-dimensional kernel feature space by using the nonlinear mapping associated with a kernel function. In the kernelized subspace, SAR target recognition is formulated as a joint covariate selection problem across a group of related tasks. Furthermore, a multi-task weight optimization scheme is developed to compensate for the heterogeneity of the multi-scale features and enhance the recognition performance. Extensive experimental results tested on the public moving and stationary target acquisition and recognition (MSTAR) dataset demonstrate that our proposed method achieves better recognition performance than other existing competitive algorithms.

Highlights

Synthetic aperture radar (SAR), operating as a special type of high resolution observation and detection system, is based on electromagnetic coherence imaging mechanism
In order to overcome this hurdle, we introduce the kernel tricks, which are well known for the ability of solving nonlinear problem by mapping the data from original space to a high dimensional space [21], into the sparse coding, so that the projected monogenic features in the feature space may become more grouped and linearly separable
We reduce the dimensionality of kernelized monogenic features by applying the kernel-based dimensionality reduction approach, and construct different sparse coding dictionaries for different monogenic features

Summary

INTRODUCTION

Synthetic aperture radar (SAR), operating as a special type of high resolution observation and detection system, is based on electromagnetic coherence imaging mechanism. The former way, corresponding to low-level (feature level) fusion, may result in loss of information; while the latter, associated with high-level (decision/score) fusion, is easy to neglect important structural information among multiple classifiers, which influences the recognition performance To solve these problems, we will develop a multi-task kernel sparse representation classifier based on an intermediate level (task level) fusion idea. Before we give its specific form, we would like to point out that the multi-task kernel sparse representation scheme has a key problem that there exist possible heterogeneities among different learning tasks These heterogeneities may come from different measuring scales or variations of the discriminative capability of each model in a certain feature space. We will introduce how to compute these task weights

MULTI-TASK WEIGHTS OPTIMIZATION

Findings

CONCLUSION