Augmenting Simulations for SAR ATR Neural Network Training

A training data augmentation technique is presented that approximates the differences between measured and simulated SAR imagery. This method is applied to simulated images and a CNN is trained with them. We achieve over 95% cross-class classification using the SAMPLE dataset from AFRL, with 1% measured data in the training set. We compare this to 89.6% accuracy when the augmentation technique is not used. Our hypothesis is that, while simulations can be made to approximate the measurements very closely, further augmentation can increase accuracy over non-augmented simulations.