Segmentation and recognition of off-line cursive handwriting.
Berrin A. Yanikoglu · 1993
Off-line handwriting recognition is the task of interpreting the image of handwritten text. In this thesis, we present a system for recognizing off-line, cursive, English text, guided in part by global characteristics (style) of the handwriting. We characterize the style by various parameters and present algorithms to extract these parameters for each text line, after locating the test line boundaries. Segmenting words into letters has been a major obstacle in cursive handwriting recognition. We introduce a new method for segmentation, based on minimizing a cost function. The cost of segmenting at a point depends on local properties of the text around that point and on the style parameters. Cursive script segmentation is inherently ambiguous since some letters and digraphs can be indistinguishable in the image. The segmentation algorithm tries to find all possible letter boundaries and as few additional ones as possible to avoid making early wrong commitments. The segmentation ambiguity is indirectly arbitrated in the word recognition phase, where each possible segmentation is evaluated to find the most likely word. After segmentation, the letters (segments) are normalized to remove size and slant variations. The novel size normalization algorithm uses the style parameters to scale different parts of a letter separately, and thus removes much of the variation in the writing. The normalized letters are then classified by a feedforward neural network. The system then recognizes the word, using the output of the neural network as posterior probabilities of letters. We compared two hypotheses for finding the likelihood of words that are in the lexicon and found that using a Hidden Markov Model (HMM) of English is less successful than assuming independence among the letters of a word, for the limited test words used in the experiments. We also used the Viterbi algorithm to find the most likely letter string, without constraining it to be in the lexicon. In our experiments with several writers, 96% of all the test words were correctly segmented, 52% of them were correctly recognized, and 70% were in the top three choices. The letter recognition performance was 55% for the letters of these words.