Comparative implementation of two fusion schemes for multiple complementary FLIR imagery classifiers

Several classifiers for forward looking infra-red imagery are designed and implemented, and their relative performance is benchmarked on 2545 images belonging to 8 different ship classes, from which 11 attributes are extracted. These are a Bayes classifier, a Dempster–Shafer classifier ensemble in which specialized classifiers are optimized to return a single ship class, a k-nearest neighbor classifier, and an optimized neural net classifier. Two different methods are then studied to fuse the results of selected subsets of these classifiers. The first method consists of using the outputs of various classifiers as inputs to a second neural net fuser. The second method consists of converting the outputs of these classifiers into masses for use in a Dempster–Shafer fuser. In both approaches, the fused classifier achieves better results than the best classifier for any given class.