Application of Threshold Techniques for Readability Improvement of Jawi Historical Manuscript Images

TL;DR

This study evaluates thresholding techniques to enhance the readability of degraded Jawi manuscript images, finding local methods like Niblack and Sauvola improve clarity, with Niblack performing best in recognition tasks.

Contribution

It introduces an effective application of local thresholding methods, especially Niblack, for improving readability of historical Jawi manuscripts with various quality issues.

Findings

Local thresholding methods outperform global techniques.

Niblack's method yields better recognition accuracy.

Enhanced post-processing can further improve results.

Abstract

Historical documents such as old books and manuscripts have a high aesthetic value and highly appreciated. Unfortunately, there are some documents cannot be read due to quality problems like faded paper, ink expand, uneven colour tone, torn paper and other elements disruption such as the existence of small spots. The study aims to produce a copy of manuscript that shows clear wordings so they can easily be read and the copy can also be displayed for visitors. 16 samples of Jawi historical manuscript with different quality problems were obtained from The Royal Museum of Pahang, Malaysia. We applied three binarization techniques; Otsu's method represents global threshold technique; Sauvola and Niblack method which are categorized as local threshold techniques. We compared the binarized images with the original manuscript to be visually inspected by the museum's curator. The unclear features were marked and analyzed. Most of the examined images show that with optimal parameters and effective pre processing technique, local thresholding methods are work well compare with the other one. Niblack's and Sauvola's techniques seem to be the suitable approaches for these types of images. Most of binarized images with these two methods show improvement for readability and character recognition. For this research, even the differences of image result were hard to be distinguished by human capabilities, after comparing the time cost and overall achievement rate of recognized symbols, Niblack's method is performing better than Sauvola's. We could improve the post processing step by adding edge detection techniques and further enhanced by an innovative image refinement technique and a formulation of a class proper method.