Statistical approach to feature extraction for numeral recognition from degraded documents

Proposes a fast statistical method for numeral recognition from degraded documents. Our method uses a feature-based approach in combination with weights assigned to each feature along with a factor that defines how much of the feature is detectable. We describe an object by how it looks or does not look by giving positive or negative weights for those features. In the proposed method, the positive and negative weights together describe a numeral completely, and help to clearly distinguish one numeral from another. The recognition rate obtained by this method is much higher than that obtained by a simple feature-based method. We present results of 160 bills printed on a dot matrix printer. There are many variations in the printed output, such as faded bills or highly smudged outputs. With this level of distortion, the numerals are sometimes ambiguous to human eyes. However our method can recognize the distorted numerals with a success rate of 97%. The algorithm takes about 0.01 seconds for the recognition of a single numeral on a Pentium 200 MHz machine.