Abstract:
This work proposes a new recognition strategy involving an efficient recognition system for Arabic manuscript. Especially, for the handwritten Arabic literal amounts that are used in the check bank. Three main contributions are proposed in the development of our recognition model. These contributions are based on the natural segmentation of the candidate word in which a set of the Part of Arabic Words PAWs is generated. Accordingly, these proposed contributions have the ability to enhance both recognitions steps, which are: the feature extraction and the classification steps. The first contribution is proposed in the feature extraction step where the structural features are extracted from the PAWs of the candidate word. Notice that, some candidate words may be provided the same structural features. To avoid this problem, the statistical features are extracted from the candidate word and then combined with the above-mentioned structural ones. The second contribution is proposed in the classification step, in which the multi-classifiers based SVMs is used instead the single classifier SMV. Therefore, the multi-classes problem may be divided according the number of the PAWs where each one is independently solved using four separate set of the classifier. The main goal of the third contribution is to solve the touching problem of certain PAWs, which is appeared when a candidate manuscript is ill-written. This problem is solved by including an additional pre-classification step where the erroneous number of the PAWs is well corrected. Finally, the proposed recognition system is validated on the AHDB data base and the obtained results are compared by those given by the conventional recognition systems. Accordingly, the proposed recognition system provides better results in terms of classification accuracy and recognition rate.
