Style Quantification of Scanned Multi-source Digits

The co-occurring patterns in a group carrying the traits of common origin are statistically dependent via an underlying style context. Exploiting style consistency in groups of patterns from multiple sources can increase OCR accuracy. The accuracy gains obtained by a style consistent classifier depend on the amount of style in isogenous (same-source) fields. We present mathematical models to quantify the amount of single-class and multi-class style using entropy, correlation and mutual information. We also demonstrate a method for style homogenization that allows testing our metrics on real data