Kernel Joint Sparse Representation Based SAR Automatic Target Recognition

Synthetic aperture radar (SAR) automatic target recognition (ATR) is an important task in computer vision. This paper proposes a novel SAR ATR method based on kernel joint sparse representation. First, a monogenic feature extraction method is adopted to capture multiscale spatial and spectral properties of targets. Second, a kernel joint sparse representation classifier (KJSRC) is designed. In this KJSRC, to make the data linearly separable and more discriminative, we integrate the kernel principal component analysis and kernel Fisher discriminant analysis to produce an augmented pseudo-transformation matrix, which maps the features from the input space to a reproducing kernel Hilbert space. Third, different from single task learning, our method builds different linear regression models for each kernel-mapped monogenic feature. Through optimizing ℓ
₁
/ℓ
_q
-norm regularization problem, the representation coefficients across multitask representation models are estimated. Finally, the query label is estimated according to total reconstruction error minimization rule from all tasks. Experimental results show that our method achieves high recognition accuracy for SAR ATR.