Segmentation of Background and Foreground for Ancient Lanna Archaic from Palm Leaf Manuscripts using Deep Learning

Abstract

This research aimed to present a method of image segmentation of background and foreground in the image processing of ancient Lanna archaic that have been recorded on palm leaf manuscripts. The proposed method began by randomising 2,500 pixels from 10 palm leaf manuscript images in a RGB colour space, followed by clustering the randomised pixels (trained dataset) using K-means, when k=(2,3,4,5,6,7, 8). Each k cluster applied was required to provide the accuracy of more than 85 percent. A deep learning algorithm was selected as a predictive model. Then, the results from the segmentation of background and foreground of archaic were compared with Otsu's method, a widely renowned image segmentation method. Consequently, the proposed method did not result in satisfactory background and foreground segmentation. When analysing the causes of inadequate results, it was found that it could be due to the inadequacy of 2,500 pixels as dataset and the similarities of 10 sampled palm leaf manuscript images in terms of light intensity and colour, whereas the palm leaf manuscripts used for verifying the proposed method were different from the trained dataset in terms of colour intensity and light intensity, hence resulting in the inefficient image segmentation.