Probabilistic Model for Segmentation Based Word Recognition with Lexicon

The problem of off-line reading of unconstrained handwritten words has been studied extensively due to its role in many important applications such as reading addresses on mail-pieces [3, 6, 11], reading amounts on bank checks [7, 10], extracting census data on forms [2, 9], and reading address blocks on tax forms [12]. The main challenges are wide variety of writing styles, poor image quality and missing or extraneous strokes caused by segmentation errors. The intuitive solution to the problem is to segment the word image into probable character sub-images, then try to recognize separate characters and combine results [4, 8]. The function of optical character recognizer (OCR) used is to provide confidence scores for supposed character images. Although many different OCRs are available, they are all mainly focused on classifying isolated character images. In practice, when dealing with unconstrained handwritten word images there is no guarantee that segmented sub-images will be single isolated characters. So OCR used for word recognition should be able to provide low confidence scores for non-character images. Besides choosing the right OCR for word recognition, it is also important to know how to incorporate OCR confidence scores for individual characters into an overall confidence score for the entire lexicon word should we take arithmetic mean, geometric mean or some other normalizing formula? This question addresses what OCR score truly means. For example, given an image and a hypothesis character should the OCR produce a score representing the posterior probability or the prior probability ? In this paper we describe the construction of a possible mathematical model for word recognizers that are based on the segmentation paradigm and use of a lexicon. The construction of the model is motivated by the comparison of two word recognizers existing in CEDAR: CMR ( Character Model Recognizer )[4] and WMR ( Word Model Recognizer )[8]. These recognizers use similar preprocessing and segmentation techniques. Using seemingly inferior character recognizer, WMR is able to perform better than CMR on word images. CMR uses the GSC (Gradient, Structural, Concavity)[5] character recognizer which is widely accepted as being very accurate.