A Complete OCR System Development of Tamil Magazine Documents

We present an early version of a complete Optical Character Recognition (OCR) system for Tamil magazine documents. All the standard elements of OCR process like deskewing, preprocessing, segmentation, character recognition and reconstruction are implemented. Experience with OCR problems teaches that for most subtasks involved in OCR, there is no single technique that gives perfect results for every type of document image. We have used the ability of artificial neural networks to learn arbitrary input/output mappings from sample data for solving the key problems of segmentation and character recognition. Text segmentation of Tamil newsprint poses a new challenge owing to its italic-like font type; problems that arise in recognition of touching and close characters are discussed. Character recognition efficiency varied from 94 to 97% for this particular type of font. The grouping of blocks into logical units and the determination of read order within each logical unit helped us in reconstructing automatically the document image in editable format. Final reconstruction is done in HTML format.