A Word Matching Algorithm in Handwritten Arabic Recognition Using Multiple-Sequence Weighted Edit Distances

No satisfactory solutions are yet available for the offline recognition of handwritten cursive words, including the words of Arabic text. Word matching algorithms can greatly improve the OCR output when recognizing words of known and limited vocabulary. This paper describes the word matching algorithm used in the JU-OCR2 optical character recognition system of handwritten Arabic words. This system achieves state-of-the-art accuracy through multiple techniques including an efficient word matching algorithm. This algorithm reduces the average sequence error for the IfN/ENIT database of handwritten Arabic words from 32.3% to an average word error of just 5.0%. This algorithm is a weighted version of the edit distance algorithm. The weighted version has a 5.0% advantage over the plain edit distance algorithm. This algorithm selects the best match utilizing a set of multiple probable sequences from the sequence transcription stage. Using multiple sequences, instead of one, reduces the average error by 27.0% over the weighted edit distance algorithm. Compared with an algorithm used in a leading system, this algorithm offers 6.7% lower average word error for the main two test sets.