Image-based ATR utilizing adaptive clutter filter detection, LLRT classification, and Volterra fusion with application to side-looking sonar

摘要: An improved automatic target recognition (ATR) processing string has been developed. The overall string consists of pre-processing, subimage adaptive clutter filtering, detection, feature extraction, optimal subset feature selection, orthogonalization and classification blocks. objects that are classified by three distinct ATR strings fused using the confidence values their expansions as features, "summing" or log-likelihood-ratio-test (LLRT) based fusion rules. These ATR were individually developed and tuned researchers from different companies. utility was demonstrated with an extensive side-looking sonar dataset. In this paper we describe a new improvement: six additional features extracted, primarily shadow information extraction window whose length is now made variable function range. This improvement resulted in a 3:1 reduction false alarms. Two advanced algorithms subsequently applied: First, nonlinear Volterra expansion (2nd order) feature-LLRT algorithm employed. Second, repeated application Volterra feature selection / LLRT block utilized. It shown cascaded Volterra feature- LLRT outperforms baseline single-stage feature-LLRT fusion algorithms, yielding significant improvements over best single results, providing the capability to correctly call majority targets while maintaining very low alarm rate.