A parallel recognition system for Arabic cursive words with neural learning capabilities

Recognizing multi-font Arabic texts is a difficult task in the area of optical character recognition (OCR) because Arabic is a cursive type language. The first chapter presents some of the work done in this field. The scope of the surveyed work varies with respect to complexity from printed isolated characters to handwritten words.; In this work, a new system for recognizing multi-font Arabic texts is proposed. Efficient preprocessing, segmentation, and preclassification algorithms have been designed and implemented to achieve this objective.; A new segmentation algorithm that has some recognition capabilities is proposed. The segmentation of words into characters consists of two main steps. The first step segments words into subwords by assigning a different color code for each connected component of the word. The second step segments subwords into characters. It produces the segmented characters along with some information including, the character form, the character height, and the color codes of the primary and secondary parts of the character. Based on these information the Arabic character set which includes 100 character classes has been decomposed into 8 small sets. The largest set includes only 12 character classes. This preclassification technique reduces the learning time taken by the classifier and increases the recognition rate. It also enhances the parallelism so that characters in different forms can be recognized simultaneously.; Several features are examined for the purpose of selecting good features for Arabic characters. The actual classification is done using multi-layer perceptrons with back-propagation learning. This paper discusses the details of each algorithm and its performance on samples from several texts. The system shows a high accuracy of over 98%.; The high speed necessary for real time applications may be achieved with emerging parallel processors. Therefore, some of the algorithms presented in this work are mapped to parallel algorithms to be executed on parallel machines. As a case study, the MasPar, which is a 2-D mesh architecture machine, is used as our experimental tool.; Finally, two skeletonization algorithms suitable for Arabic characters are proposed. These algorithms can be used in systems that require thinning.