End-to-End Machine Learning Solution for Recognizing Handwritten Arabic Documents

The research for offline handwriting recognition (HWR) solutions for various languages has recently gained rising attention, especially in the Arabic language. This is connected to the growing necessity to digitize Arabic documents in several applications such as exploring large documents, automated sorting of express mail, editing of earlier printed documents, and bank control processing. Regrettably, notwithstanding decades of research, there is no satisfactory solution for recognizing cursive handwriting like the Arabic language because of its difficulty. This paper presents end-to-end machine learning solution by applying deep learning techniques like CRNN-BLSTM using the MADCAT dataset to accurately recognize the Arabic handwritten documents after simultaneously learning text detection, segmentation, and finally conversion to editable text. Our integrated method, resulting from integrating several distinct neural networks models, achieved high accuracy in parsing the full page, converting it to lines, and predicting the writing within each line. This approach has been evaluated using a large-scale set of Arabic handwritten documents that contains various problems that need to be addressed, the achieved character error rate is 3.96%.