Automatic target recognition in SAR images using multilinear analysis

In this paper, we investigate an approach based on multilinear analysis for synthetic aperture radar automatic target recognition (SAR/ATR). High resolution SAR images are the composite consequences of multiple factors : bearing angle, grazing angle, number of views and polarisation. Linear methods like principal component analysis (PCA) require to reshape images into a high dimensional vector. This vectorisation processing breaks natural structure and correlation in the original data set. Moreover the PCA, based on a matrix singular value decomposition (SVD), allows only one factor to vary in the image database. Multilinear analysis provides a powerful mathematical framework to analyse ensembles of images resulting from the interaction of underlying factors and preserves their original shapes. In this paper, we propose a method based on multilinear principal component analysis (MPCA) to classify unlabeled targets. We form a tensor with the images of the training set and use the higher order singular value decomposition (HOSVD) to reveal interesting patterns and dependencies between images. HOSVD is also used to compress the data and remove all information belonging to the background. A multilinear projection algorithm projects the unknown target into multiple basis which characterize learned classes. Tests using real SAR images database show that the multilinear approach provides very good recognition performance with a very high compression rate.