Given large number of words to be recognized, a two-stage strategy for eliminating unlikely candidates before recognition can be a reasonable and powerful approach for increasing the recognition speed. A holistic lexicon reduction technique for offline handwritten Arabic word recognition is proposed in this paper. The principle of this technique involves the extraction of dots and subwords from the cursive Arabic word image to describe its shape. In the first stage of reduction, the number of subwords in the input word is estimated. Then in the second stage, the word descriptor, based on the dots information, is used while taking into account only the candidates selected in the first stage. Experimental results on IFN/ENIT database, consisting of 26,459 cursive Arabic word images, show a lexicon reduction of 92.5% with accuracy of 74%.
Handwritten Arabic numeral recognition using deep learning neural networks
