InkLink: A writer-dependent on-line unconstrained handwriting recognition system

InkLink is a new recognition method for on-line handwriting. Detecting poly-gram matches between words at lexically predicted locations avoids the segmentation and vocabulary limitations of character-level and word-level recognition systems, respectively. At the signal level, ligatures facilitate matching longer segments because individual characters are often indistinct. The InkLink prototype word recognition system has access to a lexicon of plausible labels, and to a set of handwritten reference words represented by extremal-point and chain code features. Two additional feature representations are obtained by reordering and pruning the time-sorted extremal-point features. For each set of features, the average number of features per letter segment is obtained from the reference set by least squares estimation.; The system consists of three stages: lexical processing, signal matching and classification. The lexical stage pre-computes the match length and location of every word in the lexicon by matching its label to the label of every reference word. The unknown word is hypothesized to be each word in the lexicon. The length and location of feature-level matches are predicted by applying the character feature-length estimates to the lexical matches between the candidate word and reference set. At these predicated locations, the signal matching stage detects the longest observed feature-level match between the unknown and each reference word. The distribution of observed feature-level match lengths, conditioned on predicted match lengths, is estimated using localized left and right word alignments. The Bayesian classification stage assigns to the unknown word the label of the lexical candidate that maximizes the probability of the observed feature-level match lengths at the predicted locations. The rank orders based on the four feature sets are combined by Borda Count.; On a typical writer, with a lexicon of 1000 words and a reference set of 1000 words, the final accuracy is 84.1%. With a lexicon of 100 words and writer-specific reference sets of 500 words, the accuracy on a dozen new writers ranges from 48.3% to 97.3%. The higher accuracies are obtained on smooth, unslanted writing. Other possible applications include handwritten and printed text in other alphabetic scripts, and speech recognition based on phonetically labeled reference sets.