Comparison of handprinted digit classifiers
1993
We report recognition results for several pattern classifiers trained and tested on disjoint sets of 30620 digits selected from the first 500 writers of NIST Special Database 3. The classifiers are ubiquitous in traditional pattern recognition literature (minimum distance, maximum a posteriori, nearest neighbor) as well as neural network literature (multilayer perceptron, radial basis functions, probabilistic neural network).For the purpose of valid comparison of classifiers fixed sets of Karhunen-Loeve Transforms, were used as features.These were produced from images preprocessed using the fixed methods for size and orientation normalization.The "K- means" clustering algorithm is used to produce subclasses thereby supervising training and aiding recognition.Graphical displays of classification and associated confidences illustrate classifier complexity.Recognition error rates for all the classifiers are tabulated as a function of feature vector dimension.Computational and memory requirements of the different classifiers are also compared.1