| Many successful pattern recognition solutions have been developed in recent years, despite the fact that pattern recognition is ill-defined and difficult due to noise and large variations in the input data. A promising approach is to fuse the decision capabilities of several classifier designs, such that they reduce the combined recognition error rate. This thesis confronts the recognition problem and proposes a method of training a Linear Fusion Network (LFN) using the Output-Reset (OR) algorithm. Training algorithms are enhanced by a closed form solution to OR. A linear network fuses the output of three types of discriminant-based classifiers: (1) Multilayer Perceptron (MLP), (2) Nearest Neighbor Classifier (NNC), and (3) Radial Basis Function (RBF) network. This framework is then applied to the task of recognition of handprinted numeral data, geometric shape data, and remote sensing data. Experimental results using OR training are compared against Minimum Classification Error (MCE) objectives. |
