A neuro-heuristic approach for segmenting handwritten Arabic text. (c2001)

Abstract

The segmentation and recognition of Arabic handwritten text has been an area of great interest in the past few years. However, a small number of research papers and reports have been published in this area. There are several major problems with Arabic handwritten text processing: Arabic is written cursively and many external objects are used such as dots, 'HanlZa', 'Madda', and diacritic objects. In addition, Arabic characters have more than one shape according to their position inside a word. More than one character can also share the same horizontal space, creating vertically overlapping connected or disconnected blocks of characters. This makes the problem of segmentation of Arabic text into characters, and their classification even more difficult. In this work a technique is presented that segments difficult handwritten Arabic text. A conventional algorithm is used for the initial segmentation of the text into connected blocks of characters. The algorithm then generates pre-segmentation points for these blocks. A neural network is subsequently used to verify the accuracy of these segmentation points. Another conventional algorithm uses the verified segmentation points and segments the connected blocks of characters. These characters can then be used as input to another neural network for classification. Two major problems were encountered in the above scenario. First, the segmentation phase proved to be successful in vertical segmentation of connected blocks of characters. However, it couldn't segment characters that were overlapping horizontally, and this affects any neural network classifier. Second, there are a lot of handwritten characters that can be segmented and classified into two or more different classes depending on whether you look at them separately, or in a word, or even in a sentence. In other words, character segmentation and classification, especially handwritten Arabic characters, depends largely on contextual information, and not only on topographic features extracted from these characters.