A probabilistic method for keyword retrieval in handwritten document images

Keyword retrieval in handwritten document images (word spotting) is very challenging given that OCR accuracy is not yet adequate for handwritten scripts, specially with large lexicons. Various proposed approaches build indices on information such as image features or OCR scores and have improved the performance of the traditional approach that builds index on OCR’ed text. In this paper, we improve existing keyword retrieval (word spotting) method by modeling imperfect word segmentation as probabilities and integrating the word segmentation probabilities into the word spotting algorithm. The word recognition scores are also converted into probabilities that are compatible with the probabilistic word spotting model.