A hybrid deep learning model to recognize handwritten characters in ancient documents in Devanagari and Maithili scripts

Any optical character recognition (OCR) system recognizes each and every character present in any document image. But, the task of performing OCR in ancient handwritten document images is challenging due to the existence of faded text and dark spots in the ancient document images. The presence of intrinsic patterns of characters and large number of character classes in most of the Indian scripts make this task even more challenging. This research article proposes a novel hybrid deep learning based OCR method to recognize each character present in ancient handwritten document image written in two different Indian scripts, Devanagari and Maithili. Various discriminating features have been extracted from each character present in the document using several convolutional layers and each extracted feature vector has been classified into a proper character class using hybrid deep learning model. The hybrid deep learning model consists of one dense layer and several recurrently connected hidden layers. Both long-short-term-memory (LSTM) and Bidirectional-long-short-term-memory (Bi-LSTM) variants of recurrent neural network (RNN) have been employed in the portion of recurrently connected hidden layers of hybrid deep learning model. The performance of the proposed OCR method has been evaluated on two self-generated datasets of ancient handwritten document images in Devanagari and Maithili scripts. The proposed method has achieved the character recognition accuracy of 96.97 percent and 95.83 percent in Devanagari and Maithili scripts respectively. The experimental results demonstrate that the proposed OCR method outperforms the state-of-the-art methods in this regard.