EMSI has considerable experience developing systems that exploit the intersecting fields of deep learning and radar phenomenology. In early 2015 we began applying deep nets to radar imagery using Theano, a Python Deep Learning Framework, and since then have stayed at the bleeding edge by adapting commercial/academic published networks to more appropriately suit the RF domain. We refuse to dismiss the physics of the problem, and actively incorporate this intuition into all aspects of our pipelines and network architectures.
Our first application of convolutional networks to the publicly available MSTAR dataset offered insight on the well-known propensity of these models to overfit to background clutter. The MSTAR dataset is an extensive collection of synthetic aperture radar images of military vehicles; the train set is an image collection of the vehicles at 17 degree elevation angle and the test set is at 15. All the vehicles stayed unmoved during the time between the two different collects, which means the background clutter was nearly the same. This raised the interesting problem of trying to develop a learned model that could take advantage of shadow information while still maintaining an appropriate level of invariance to the background clutter, which are seemingly contradictory goals.
As a detailed and extensive experiment, the MSTAR collects included multiple vehicles of the same class in the dataset, namely the BMP2 and the T72 class. This enabled a more appropriate way to validate model performance, but still allowed for overfitting to the overall operating conditions, namely a relatively homogenous rural clutter background. Going forward, EMSI’s main focus is extending SAR exploitation to highly unconstrained operating conditions using machine learned models that understand the high level of real-world variation in image data.
EMSI has developed technology to train deep learning CNNs with synthetically generated data. Our CNNs have been demonstrated to generalize well from the synthetic training data to operational test imagery.