A fuzzy logic approach to detector scoring

How do we judge the goodness of a new pattern recognition technique? The standard approach is to test it on labeled samples. This assumes that we have accurately labeled data. In many imaging applications, that assumption is not so easy to make. In particular, in an Automatic Target Detection system that produces “bounding boxes” for LADAR range images, there is considerable uncertainty deciding if the boxes determined by the algorithm actually correspond to a hit. An improved bounding box matching algorithm for detector scoring on synthetic data is presented. The first modified version uses fuzzy intersection and union operators to combine membership values from multiple matching criteria. The second version uses a fuzzy knowledge base using the compositional rule of inference with rules generated symmetrically from worth functions. Previous mistakes of crisp matching criteria are shown and the improved results of the two new methods are discussed