A New Framework for Recognition of Heavily Degraded Characters in Historical Typewritten Documents Based on Semi-Supervised Clustering

This paper presents a new semi-supervised clustering framework to the recognition of heavily degraded characters in historical typewritten documents, where off-the-shelf OCR typically fails. The constraints are generated using typographical (collection-independent) domain knowledge and are used to guide both sample (glyph set) partitioning and metric learning. Experimental results using simple features provide encouraging evidence that this approach can lead to significantly improved clustering results compared to simple K-Means clustering, as well as to clustering using a state-of-the art OCR engine.