Collaborative Combination of Neuron-Linguistic Classifiers for Large Arabic Word vocabulary Recognition

Most of the actual research in writing recognition focuses on speci ̄c applications where the vocabulary is relatively small. Many applications can be opened up when handling with large vocabulary. In this paper, we studied the classi ̄er collaboration interest for the recognition of a large vocabulary of arabic words. The proposed approach is based on three classi ̄ers, named Transparent Neuronal Networks (TNN), which exploit the morphological aspect of the Arabic word and collaborate for a better word recognition. We focused on decomposable words which are derived from healthy tri-consonantal roots and easy to proof the decomposition. To perform word recognition, the system extracts a set of global structural features. Then it learns and recognizes roots, schemes and conjugation elements that compose the word. To help the recognition, some local perceptual information is used in case of ambiguities. This interaction between global recognition and local checking makes easier the recognition of complex scripts as Arabic. Several experiments have been performed using a vocabulary of 5757 words, organized in a corpus of more than 17 200 samples. In order to validate our approach and to compare the proposed system with systems reported in ICDAR 2011 competition, extensive experiments were conducted using the Arabic Printed Text Image (APTI) database. The best recognition performances achieved by our system have shown very promising results.