Malayalam Character Recognition from Palm Leaves Using Deep-Learning
Remya Sivan, Tripty Singh, Peeta Basa Pati
Abstract
Ancient manuscripts like palm leaves, available in museum libraries, are a rich source of knowledge. Digitization helps store this knowledge protected for the future & enables its global access. Varying writing styles, presence of currently discarded & rare characters, quality of imaging, and palm leaves are some of the challenges to be handled while building an offline handwritten recognition system for these manuscripts. This paper focuses on recognizing Malayalam characters available in palm leaves using deep learning techniques. With the help of the histogram and contour method, lines are segmented from palm leaves first. Subsequently, individual characters are extracted from the lines. A customized Convolution Neural Network (CNN) is employed to recognize these segmented characters. This trained CNN recognizes forty-eight classes of segmented characters with 86% accuracy. Additionally, this paper compares the results with other standard CNN models.