Segmentation and recognition of calligraphic text

Calligraphic Hebrew text was scanned using a conventional scanner yielding a facsimile type of image. This image was aligned into lines of symbols by taking pixel density histograms in many directions. Separation of overlapping lines and symbols was made using an edge detecting borderline follower technique. After finding and circumscribing the first symbol, a search was made for expected nearby islands, such as strokes and points, before searching horizontally for the next symbol. The size of Hebrew symbols may vary within wide limits, because they are stretched to fill a line, instead of leaving a blank. Identification therefore requires the use of features which are independent of size and stretching. A tolerance to different pen widths is also required so that the normally used template pattern matching algorithms did not appear to be applicable. Recognition is carried out with the neural network. The weight of each feature is automatically adjusted during a training period with typical model text. The developed software provides a basis for a number of experiments in recognizing handwritten text, using various mixes of features. It is hoped that a universal set will be found that will give acceptable results when tested with several common types of writing. Different types of neural networks and architectures are currently under review