Sparse Signal Models for Data Augmentation in Deep Learning ATR

Automatic Target Recognition (ATR) algorithms classify a given Synthetic Aperture Radar (SAR) image into one of the known target classes using a set of training images. Recently, learning methods have shown to achieve state-of-the-art classification accuracy if abundant training data is available sampled uniformly over the classes and their poses. In this paper, we consider the problem of improving the generalization performance of learning methods in SAR-ATR when training data is limited. We propose a data augmentation approach using sparse signal models that capitalizes on commonly observed phenomenology of wide-angle synthetic aperture radar (SAR) imagery. Specifically, we exploit the sparsity of the scattering centers in the spatial domain as well as the limited persistence of the scattering coefficients in the azimuthal domain to solve the ill-posed problem of over-parametrized model fitting. Using this fitted model, we synthesize new images at poses not available in training set to augment the training data used by CNN. We validate the performance of the proposed model based data augmentation strategy on subsampled versions of the MSTAR dataset. The experimental results show that for the training data starved region, the proposed method provides a significant gain in the generalization performance of the resulting ATR algorithm.