Very good performance for the classification of handwritten digits has been achieved using feedforward backpropagation networks (LeCun et al. 1990; Martin and Pittman 1990). These initial networks were trained and tested on clean, well-segmented images. In the real world, however, images are rarely perfect, which causes problems. For example, at one time one of our best performing digit classifiers interpreted a horizontal bar as a 2; in this example the most useful response would be to reject the image as unclassifiable.
The aim of the work reported here was to train a network to reject the type of unclassifiable images (“rubbish”) typically produced by an automatic segmenter for strings of digits (e.g., zip codes), while maintaining its performance level at classifying digits, by adding images of rubbish to the training set.