A Self-Adaptive Method for Extraction of Document-Specific Alphabets

Recognition and encoding of digitized historical documents is still a challenging and difficult task. A major problem is the occurrence of unknown glyphs and symbols which might not even exist in modern alphabets. Current pre-trained OCR-methods hardly deliver usable results for such documents. This paper describes an alternative approach and framework for handling printed historical documents without restrictions on the contained alphabets or fonts. The basic idea is to derive all information required for encoding directly from the document itself. This is achieved by extracting a document-specific prototype alphabet of locatable glyphs. Core of the system is a customized clustering method which adapts automatically to new documents by ascertaining appropriate threshold parameters based on the special characteristics of glyphs. This way, the system is able to run without manual interventions and can be integrated into automated mass digitization workflows.