An Adaptive Unified Algorithm for Both Detection and Recognition

An adaptive neural-network approach to targetand clutter-modelling is introduced. A key novelty of this approach is that both targets and clutter can be modelled within the same neural network, so that detection and recognition can take place simultaneously within an integrated framework. The approach can therefore be applied across the spectrum of ATR discrimination levels, e.g.: detection of unknown targets in clutter; detection of specific designated targets in clutter; recognition of target subclass postdetection. The approach is designed to be generically applicable, to data from a variety of sensors, including HRRPs, SAR intensity imagery, complex SAR imagery, visible and EO imagery, and burstillumination LIDAR. This generic applicability is attributable to the fact that the algorithm adaptively models training-exemplar data of arbitrary type and dimensionality. Unlike many current approaches to target detection, this approach can exploit a wide range of cues for discriminating targets from clutter objects, including detailed grey-level shape information and, for RF sensors, complex/phase information. Furthermore, the approach is quick to use in operation, and has been designed with hardware implementations in mind. Successful results are presented for a target (designated building) detection and identification problem using real SAR imagery. The approach has been designed to have the future potential to offer other very significant new capabilities, e.g. the potential for reducing false-alarm rate in urban clutter and improving robustness to extended operating conditions.