Matthew Miller of University of Maryland on Developments in Arabic Script Optical Character Recognition: Technical Advances and Organizational Barriers #WIDH2018
Benjamin Kiessling’s Kraken was the start. Does not require character segmentation, which eliminates a common source of error in Arabic and Persian. A thousand lines of training data for a certain typeface was fed, eventually reaching 98.7% accuracy. #WIDH2018
Scan quality was not as important as initially hypothesized for reaching this accuracy. #WIDH2018
But was it generalizable to other typefaces? If you train a model, you can get very accurate digital texts. #WIDH2018
Worked with JSTOR to create an Arabic pilot using the publicly available al-Abhath journal (going back to 1948). Arabic-script only error rate: 98.46% #WIDH2018
For Persian: 98.62% #WIDH2018
Detailed accuracy study: persistent problems in Arabic and Persian OCR: diacritics, white space, Arabic in Persian texts, doubling of letters, multi-language failure, ligatures, skewed lines, footnotes #WIDH2018
Line segmentation is also an issue, the fix for which is under way. #WIDH2018
Collaborating with @SHARIAsource to create an interface for the Arabic OCR geared towards scholars and researchers. Includes a number of useful features, such as robust version control. #WIDH2018
To build a robust scholarly corpus: there are organizational barriers, need sustainable funding, need topic editors. #WIDH2018
Keep an eye out for updates regarding this Arabic OCR project and the accompanying interface (with @M_T_Miller @maximromanov @sarahsavant1 @SHARIAsource) #WIDH2018
