HIERARCHICAL INTEGRATION OF SENSOR DATA AND CONTEXTUAL INFORMATION FOR AUTOMATIC TARGET RECOGNITION

Real-time assessment of high-value targets is an ongoing challenge for the defense community. Many automatic target recognition (ATR) approa ches exist, each with specific advantages and limitations. An ATR syst em is presented here that integrates machine learning, expert systems, and other advanced image understanding concepts. The ATR system emplo ys a hierarchical strategy relying primarily on abductive polynomial n etworks at each level of recognition. Advanced feature extraction algo rithms are used at each level for pixel characterization and target de scription. Polynomial networks process feature data and situational in formation, providing input for subsequent levels of processing. An exp ert system coordinates individual recognition modules. Heuristic proce ssing of object likelihood estimates is also discussed. Here, separate estimators determine the likelihood that an object belongs to a parti cular class. Heuristic knowledge to resolve ambiguities that occur whe n more than one class appears likely is discussed. In addition, a comp arison of model-based recognition with the primary polynomial network approach is presented. Model-based recognition is a goal-driven approa ch that compares a representation of the unknown target to a reference library of known targets. Each approach has advantages and limitation s that should be considered for a specific implementation. This ATR ap proach can potentially overcome limitations of current systems such as catastrophic degradation during unanticipated operating conditions, w hile meeting strict processing requirements. These benefits result fro m implementation of robust feature extraction algorithms that do not t ake explicit advantage of peculiar characteristics of the sensor image ry; and the compact, real-time processing capability provided by abduc tive polynomial networks.