Word detection applied to images of ancient Roman coins

This paper presents a method for recognizing legends in images of ancient coins. It accounts for the special challenging conditions of ancient coins and thus does not rely on character segmentation contrary to traditional Optical Character Recognition (OCR) methods designed for text written on paper. Instead, characters are detected by means of individual character classifiers applied to a dense grid of local SIFT features. Final word recognition is accomplished using a lexicon of known legend words. For this purpose, the Pictorial Structures approach is adopted to find the most likely word occurrences based on the previously detected characters. Experiments are conducted on a set of 180 coin images from the Roman period with 35 different legend words. Depending on the lexicon size used, the achieved word detection rate varies from 29% to 53%.