Devanagari Ancient Manuscript Recognition Using AlexNet

Ancient manuscripts are one of the major source of obtaining valuable information. Reading these manuscripts is a tedious task so there came a requirement of digitization of these images. Many such OCRs exist which can decode the manuscripts and convert them into a human readable form. One such OCR is developed in this paper which was able to read regional language-Devanagari. For this classification AlexNet transfer learning model is used. In this paper Devanagari manuscript data set is used to classify the Devanagari characters. A total of 7356 characters were included in the dataset which was a combination of Devanagari basic characters and some modifiers and conjuncts. The experiments were performed after dividing the dataset into 90–10 (P1), 80–20 (P2) and 70-30(P3) train-test data. A maximum accuracy of 95.4% was obtained when the model was made to run on P1 partitioning strategy on 30 epochs. The average accuracy of 93.3% obtained for AlexNet model. Also, this paper compares the performance of other CNN based OCRs.