A novel Algorithm for Persian OCR
کد مقاله : 1059-ISME
نویسندگان:
امیر غفاری *1، علی غفاری2، علی هل اتایی3
1دانشجوی دانشگاه صنعتی خواجه نصیر الدین طوسی - رشته مکانیک
2استاد گروه طراحی کاربردی - رشته مکانیک دانشکده مکانیک دانشگاه صنعتی خواجه نصیر الدین طوسی
3مکاترونیک دانشکده مکانیک دانشگاه خواجه نصیر الدین طوسی - تهران - ایران
چکیده مقاله:
Abstract—Developing optical character recognition systems for cursive languages like Persian and Arabic has been a challenging task due to their complex structures and frequent connected letters. In this paper a new algorithm is proposed for Persian OCR using a combination of convolutional and recurrent neural networks (CNN and RNN) and CTC as our loss function. The proposed network is trained with the end – to – end fashion. Thus, all parameters are trained jointly and no pre- processing or post- processing action is required. This algorithm could overcome the variety and complexity of Persian notations. Comparison between the results of our proposal and the well-known Google Doc. OCR engine shows that for the normal Persian fonts, we have recognized 95.9 percent of words versus 93.2% of Google. For Nastaliq fonts, the proposed algorithm recognizes 96.97% of characters and 87.1% of words, while Google has not yet presented OCR for Persian Nastaliq.
کلیدواژه ها:
Persian OCR, CRNN, Nasir OCR Dataset
وضعیت : مقاله برای ارائه شفاهی پذیرفته شده است