Optimizing error-reject trade off in recognition systems

This paper describes an approach to design decision making modules in recognition systems. The input of a decision maker is a list of possible alternative decisions ordered according to their scores. The decision making task is interpreted as distinguishing "good" lists, where the correct decision has the best score and is on the top of the list, from all other lists. This is a two-class recognition problem, to solve which we define a feature set and use a neural network recognizer. The neural network estimates posterior probabilities of classes, so it is possible to make optimal (Bayes) decisions by comparing the probability of "good" list class with a single threshold. By changing this threshold and measuring error/rejection rate on a test set, one can estimate the error-reject trade off of the designed decision maker. Implementation of the approach in the A2iA bank check recognition system as well as experimental results are presented.