DEVELOPMENT OF FAST-TIMING 3-D PHOTODETECTORS FOR PET

Not Available For every Automatic Target Recognition (ATR) application, a class of clutter can be defined as target-like because most false alarms originate from that category. If this class of target-like clutter can be accurately defined for a given ATR application and precomputed, then the false alarm rate for that application can be significantly reduced and throughput rate increased. In particular, the classifier's decision surface can be crafted much more accurately through the use of negative training examples (training samples labeled non-target ). The first step of our proposed approach is to cluster all clutter in a self-organized manner using a hierarchical Adaptive Resonance Theory neural network. Self-organization is critical since only an unbiased analysis across the full clutter space will allow us to uncover all potential target-like clutter clusters. The second step is to analyze each cluster using the statistical classification technique known as Support Vector Machines to choose members from the clutter categories to use as negative training examples for the ATR's own classifier. This approach enables automatic selection of the members ( support vectors ) closest to the target category in feature space, those members most likely to cause false alarms. The approach will be evaluated on in-house developed ATR algorithms.