Sparse Signal Models for Data Augmentation in Deep
 Learning ATR

Automatic target recognition (ATR) algorithms are used to classify a given synthetic aperture radar (SAR) image into one of the known classes by using information gleaned from set training images that available for each class. Recently, deep learning methods have been shown achieve state-of-the-art classification accuracy if abundant data available, especially they sampled uniformly over and in their poses. In this paper, we consider ATR problem when limited available. We propose data-augmentation approach incorporate SAR domain knowledge improve generalization power data-intensive algorithm, such as convolutional neural network (CNN). The proposed method employs physics-inspired limited-persistence sparse modeling approach, which capitalizes on commonly observed characteristics wide-angle imagery. Specifically, fit over-parametrized models scattering data, use estimated synthesize new at poses sub-pixel translations not order augment data. exploit sparsity centers spatial smoothly varying structure coefficients azimuthal solve ill-posed model fitting. experimental results show that, data-starved regions, provides significant gains resulting algorithm’s performance.