A comparative study of combination schemes for an ensemble of digit recognition neural networks

Many researchers have been concerned with classifier combination techniques for complex pattern recognition problems such as digit recognition. Several combination approaches are compared with respect to achieving very high accuracy digit recognition. Both nonadaptive and adaptive combining schemes were considered. Non adaptive combiners studied included: majority voting, highest confidence, Borda count, and vector addition. Adaptive combination schemes investigated included: four different neural net combiners and weighted voting optimized with a genetic algorithm. These different schemes were tested on the task of combining three neural network digit classifiers, identical in architecture but trained using different training sets. Significant increases in recognition performance on blind test sets containing non digit blobs are reported in the region of very high accuracy (0.2% to. 1.0% error) for the neural network and genetic algorithm approaches on a database of digits extracted from financial documents