Out-of-Library SAR Target Recognition with Deep Learning from Synthetic Data and Multiview Information Fusion

In open-fields, a properly designed neural network for SAR automatic target recognition (ATR) should be able to produce super-class labels for targets it has never seen before, i.e. out-of-library (OOL) targets. Moreover, the network’s generalization ability should be tested with SAR imagery collected by multiple types of SAR systems. Unfortunately, most of the existing SAR ATR algorithms have only been tested by the Moving Stationary Target Acquisition and Recognition (MSTAR) dataset due to the lack of other publicly available data, which leads to potentially biased results. By jointly exploiting the Synthetic and Measured Paired Labeled Experiment (SAMPLE) dataset recently public-released by the U.S. Air Force Research Laboratory (AFRL) and the proprietary NUAA mini-SAR dataset, the problem of super-class labeling for OOL targets is thoroughly investigated for the first time. Simulation results show that super-class prediction for OOL targets with 100% synthesized SAR imagery as training data is technically feasible. Moreover, we demonstrate that fusing the information from the SAR images acquired at different elevation angles would boost the performance of the ATR algorithms for non-ideal SAR image samples.