U-NET: CONVOLUTIONAL NEURAL NETWORK FOR BINARIZATION OF HISTORICAL KANNADA HANDWRITTEN PALM LEAF MANUSCRIPTS

Ancient documents are significant historical treasures that hold valuable information about past cultures and civilizations.Binarization ofhistorical Palm leaf'smanuscriptis a tedious task, as they contain debris, noise and degradednessin nature.The existing binarization algorithms may not effectively remove all types of noise present in the documents.Recently, growing interest has been shown in creating image approaches applying deep learning models as a result of their success in numerous vision applications.In this study, we propose a U-Net architecture which is part of Convolutional Neural Network (CNN) for the binarization of ancient Kannada handwritten palm leaf's manuscript.The proposed method of CNNs to learn complex image representations and handle the variability and complexity of the palm leaf.The performance estimation is measured by calculating Precision, Recall, F-Measure, MSE&PSNR and validating them with manual results obtained by language specialistsandepigraphist's.Additionally, the results are also compared with other standard methods, such as Sauvola and Niblack, on several datasets including H-DIBCO, AMADI_LONTARSET, PHIBD-2012 and our own Historical Kannada Handwritten Palm leaf (HKHPL) dataset.Binarization of historical Kannada handwritten palm leaf manuscripts is critical for determining ageand recognizing ancient Kannada characters.The proposed method demonstrates its efficacy in these applications and highlights the potential for deep learning models to improve binarization of historical documents.