Neural-Based Segmentation Technique for Arabic Handwriting Scripts

Abstract

In some algorithms, segmentation of the word image considers the first step of the recognition processes; the main aim of this paper is proposed new fusion equations for improving the segmentation of word image. The technique that has used is divided into two phases; at the beginning, applying the Arabic Heuristic Segmenter (AHS), AHS uses the shape features of the word image, it employs three features, remove the punctuation marks (dots), ligature detection, and finally average character width, the goal of this technique is placed the Prospective Segmentation Points (PSP) in the whole parts of the word image. As a result, the second phase apply the neuralbased segmentation technique, the goal of neural technique is check and examine all PSPs in the word image in order to report which one is valid or invalid, this will increase the accuracy of the segmentation; to do that, the network obtains a fused value from three neural confidences values: 1) Segmentation Point Validation (SPV), 2) Right Character Validation (RCV), and 3) Central Character Validation (CCV) which will assess each PSP separately. The input vectors of the neural network are calculated based on Direction Feature (DF), DF considers much more suitable for Arabic Scripts. AHS and neural-based segmentation techniques have been implemented and tested by local benchmark database.