Recognition of Handwritten Hebrew One-Stroke Letters by Learning Syntactic Representations of Symbols

An automatic recognition algorithm of cursive one-stroke Hebrew letters is presented. The recognition procedure is a first but major part of a full Hebrew script recognition, as Hebrew script is naturally written in separated characters. The algorithm is founded upon representation of each character by basic structures: lines, arcs, loops, and edges arranged in a chain list. The structures are presented in normalized domains that establish position and direction, and emphasize edges and horizontal and vertical directions. For every input letter the system creates a chain list which represents it, and compares it with letters previously learned that are included in its knowledge-base (KB). If a perfect or partial match occurs between the input list and one of KB's entries the letter is recognized; otherwise the user is asked to identify the letter, and the letter representation is added to KB. Following every recognition the user tests the system response, and in case of an error the letter representation is entered into KB as a new entry. A special learning procedure is included to make possible an efficient adaptation from one handwriting to another. The system strengthens successful entries and weakens entries which cause error or are not used. The system was tested on cursive Hebrew script written by four different people. Starting with an empty KB, the system achieves an average recognition rate of about 85 percent correct, 10 percent unknown, and 5 percent error. The adaptive learning procedure makes possible a recognition rate of almost 100 percent for a given writer after each character has appeared about ten times.