Strategic defense systems need to develop Automated Target Recognition (ATR) sub-munitions that can identify friend or foe and for surveillance. Current ATR methods have problems classifying highly varying images of 3D objects. The proposed work aims to overcome some of these difficulties by building a neural network ATR system that learns by example. The objective of the proposed Phase I effort is to show proof-of-concept by classifying multiple toy airplanes in different lighting conditions and 3D orientations. The proposed effort will extend a working recognition product that has been shown to achieve extremely high accuracy in recognizing highly variable symbols. The neural classifier has been designed to use general features that are not specific to the class of objects being recognized. Its parallel architecture allows highly optimized processing speeds. It will tolerate internal noise, partial system damage and parameter drift. In Phase II of this project, the neural network ATR system will continue to be developed to demonstrate fast recognition of aircraft in the field.
