Computational Intelligence Approaches to Automatic Target Recognition

An approach is desired that allows composite infrared background scenes to be synthetically generated using a data base of background objects and types. I-MATH has previously developed an algorithm for a related application, whereby actual scene images are decomposed with a Laplacian pyramid, and then each level in the pyramid is characterized by its second order statistics. Synthetic scene images are synthesized to have the same characteristics, and the synthetic images are determined to be sufficient by a "closeness" algorithm. Before the synthetic generation of backgrounds can be accomplished, a set of background metrics must be identified, I-MATH has been performing research into such metrics for ten years and is very aware of the limitations of such popular parameterizations as the Schmeider-Wethersby metric for describing clutter. For the Schmeider-Wethersby metric, there is no representation of target shape, target internal detail, or background clutter. Hense, against even bland backgrounds, checkerboards can have the same equivalent DT as a uniform target. I-MATH originally addressed this deficiency by adapting a camouflage metric, which convolves the brightness-structured target by the point spread function of the eye (or other sensor), and then computes an equivalent DT using absolute values. In a subsequent development of a Visual Observer Model (VOM), I-MATH incorporated a target-to-background discriminability metric based on cooccurrence metrics. This computes a density of confusing forms (M) and a corresponding discrimination probability (P3). In Phase I, we will evaluate several additional M and equivalent DT metrics. A representative variety of clutter conditions will be considered, giving consideration to a recently developed taxonomy for distinguishing clutter types. The characterization is three dimensional, with the bases related to (1) repetitiveness/fractal (dimension), (2) directionality (gradient), and (3) complexity (granularity). For Phase I, we will use the existing P3 formulation in VOM for computing target discrimination probability. This is an empirically derived and well validated equation for predicting operator image assessment. In Phase II, we will repl