A probabilistic method for keyword retrieval in handwritten document images

Keyword retrieval in handwritten document images is a challenging task because handwriting recognition does not perform adequately to produce the transcriptions, specially when using large lexicons. Existing methods build indices using OCR distances or image features for the purpose of retrieval. These alternative methods are complimentary to the traditional approaches that build indices on OCR'ed text. In this paper, we describe an improvement to the existing keyword retrieval (word spotting) methods by modeling imperfect word segmentation as probabilities and integrating these probabilities into the word spotting algorithm. The scores returned by the word recognizer are also converted into probabilities and integrated into the probabilistic word spotting model.