Reject Options and Confidence Measures for kNN Classifiers
Christoph Dalitz · 2009
This survey summarizes proposals made in the pattern recognition literature for detecting uncertain patterns that should rather be rejected than classified by a classifier. Beyond reviewing methods applicable to distance based nearest neighbor classifiers, this article describes an interface for computing confidences, storing them with classified images and querying this information, as it is implemented in the Gamera framework for document analysis and recognition. Based on this interface, a method for detecting broken, touching, and unknown characters, as well as noise, is proposed. The method is applied to two historic prints, showing that this method works well for detecting broken and touching characters, but less reliable for identifying glyphs representing noise.