Release Notes

This work originally appeared in the Proceedings of the 5th USPS Advanced Technology Conference, Washington, D.C. Copyright is retained by the authors. This work may be freely copied for non-commercial purposes provided this copyright notice is included.

Abstract

Many character recognition techniques produce, in addition to the recognition result, an estimate of the confidence level of that result. The confidence estimate is used to obtain the desired reject/misread tradeoff in image recognition systems. We present methods for improving and normalizing these confidence estimates. This calibrated confidence gives a more accurate estimate of the likelihood that a recogntion results is correct, which in turn permits more accurate estimation of confidence over fields of characters (e.g. ZIP codes).

The calibration is accomplished by treating the recognizer as an unknown filter function, which distorts a hypothetical true confidence, producing an uncalibrated output confidence. The unknown filter function is estimated using supervised learning methods drawn from signal processing and neural network techniques. From this estimate, an inverse filter is constructed that approximately reverses the distortions introduced by the recognizer, thus yielding the calibrated confidence.