Training a CNN with Simulated Data for ATR - Lessons Learned

During the last decade, Convolutional Neural Networks (CNNs) have revolutionized many areas of electro-optical (EO) image processing, regularly surpassing traditional methods. This success is strongly connected with the availability of labeled data sets such as ImageNet, and the technique of transfer learning. Inspired by this rising interest, the foundations and building blocks of CNNs have been widely extended and many new techniques for optimization, better convergence or better generalization have emerged in very short time. For SAR images, though evolving, the CNN literature is much sparser and it is not entirely clear yet, which techniques from the optical domain also work for SAR images. This paper takes up the example of SAR Automatic Target Recognition (ATR), training with simulated data, and gives an overview of some of the techniques propagated in the EO literature. These were put to the test on SAR ATR on an airborne data set for vehicle classification and evaluated as to their usefulness for this task.