Efficiently Computing and/or Compensating for Object Variability in Automatic Target Recognition (ATR) Applications

A research program is proposed on the integration of signal processing and electromagnetic modeling, to address the problem of performing ATR with targets possessing a high degree of variability. Sparseness is employed from two perspectives. First, in the signal processing component, sparse classifiers are developed, based on principled Bayesian techniques, which infer the scattering physics most relevant for ATR applications. This relevant scattering phenomenology is linked to the physical components of the target, to focus computational resources. By defining the sparse set of key scattering features, one implicitly infers which relatively small set of observables are most robust to target variability, while also providing discriminative power. The second area in which sparseness is employed is within the computational electromagnetic model. Compressive sensing employs the fact that the angle-frequency dependent scattered fields are typically sparsely rendered in an orthonormal basis (wavelets or DCT), and based upon this one need only perform a relatively small number of computations, from which the remaining computations may be inferred. By exploiting the joint information across multiple similar but distinct targets in a database, the number of compressive computations may be further reduced. The proposed research seeks to optimally integrate these processing and computational tools.