![]() | Only 14 pages are availabe for public view |
Abstract The complexity of Arabic Letters lies in the fact that each let- ter in each different position (start, end, middle, and isolate) is represented with a different shape and style of writing. But the similarity occurs between different letters in different positions, which raises conflict when it comes to recognizers and classifiers. Those facts, in addition to the cursive nature of Arabic language, along with variations in writing style, methods, and fonts, makes Arabic Handwriting a fairly complex task to perform. In this research, a new approach to Handwriting recognition is introduced. The method involves the training of a separate HMM for every Arabic letter in the alphabet in each of their various positions. The method followed is based on diacritics removal, along with similar letter grouping, to reduce the total number of models to be trained to improve the overall efficiency of the system. The system also uses a set of hybrid high perfor- mance features to help the HMM identify the main characteris- tics of each letter and help identify the differences between them. And finally after HMM has performed its recognition phase, the post-processing algorithm is used to improve on the recognition iii rate of the HMM, reaching a recognition rate of around 87%. The proposed system is trained and tested using the IFN/ENIT database which contains tens of thousands of images handwritten in Arabic from different writers to provide the needed variation for improved training process and non-biased recognition. |