Asymptotic accuracy of two-class discrimination
T.K. Ho, Henry S. Baird · OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) · 1994
Poor quality-e.g. sparse or unrepresentative-training data is widely suspected to be one cause of disappointing accuracy of isolated-character classification in modern OCR machines. We conjecture that, for many trainable classification techniques, it is in fact the dominant factor affecting accuracy. To test this, we have carried out a study of the asymptotic accuracy of three dissimilar classifiers on a difficult two-character recognition problem. We state this problem precisely in terms of high-quality prototype images and an explicit model of the distribution of image defects. So stated, the problem can be represented as a stochastic source of an indefinitely long sequence of simulated images labeled with ground truth. Using this sequence, we were able to train all three classifiers to high and statistically indistinguishable asymptotic accuracies (99.9%). This result suggests that the quality of training data was the dominant factor affecting accuracy. The speed of convergence during training, as well as time/space trade-offs during recognition, differed among the classifiers.