Evolving Spatially-Localized Projection Filters for SAR Automatic Target Recognition

摘要

A new approach to synthetic aperture radar (SAR) image object modeling is proposed, with a concept of `object' consisting of a small set of spatially-localized filters. An evolutionary programming (EP) approach provides the mechanism for automated model generation and optimization. Target models are developed using actual SAR image signatures of armored military vehicles. The evolutionary programming approach is shown to provide an efficient technique for model generation, experimentally demonstrating fast convergence properties. The classification efficacy of the approach is demonstrated in a two-class environment via discrimination of T-72 tanks vs. two other physically similar armored vehicles. Promising classification results, obtained experimentally through blind testing, are presented.