Optical Character Recognition through Character-set Dependent Probabilities

While optical character recognition is field where machine learning algorithms are easily applied, it might not always be cost effective to do so. While Viola and Jones explored Haar features in their face detection algorithm, selecting from the thousands of features is quite time consuming. This paper compares their feature set with a baseline of pixel learners as well as another set of Haar features that were selected based on knowledge of the character set. While probability-chosen Haar features are a computational improvement over using Adaboost-selected Haar features, they don't perform as well as the raw pixel features, though there is room for improvement.