A Survey of Math Accessibility For Blind Persons and An Investigation on Text/Math Separation

Despite recent advances, blind students, researchers, and professionals lack easy access to mathematical resources. This lack of access is a barrier to higher education for many blind students and puts them at an unfair disadvantage in school, academia, and industry. A survey of current mathematical accessibility technologies for blind persons is covered in this paper, encompassing reading, writing, and editing mathematical equations and graphics. The current research problems in this field are studied, concentrating on mathematical OCR (optical character recognition). It can be useful to separate math from plain text in documents because math OCR systems do not perform as well as regular OCR systems on text, and regular OCR systems do not perform as well on math. Improved recognition results could help the Tactile Graphics Project (TGP) at the University of Washington as well. The TGP develops techniques to automate the production of tactile images in math and science books for blind persons. Using the TGP’s current workflow, a human translator spends a significant amount of time hand-editing the text recognized by the OCR system, so improving recognition results by dealing more intelligently with math versus text could help speed up the process. The results of InftyReader, a mathematical OCR system, are merged with those of a regular commercial OCR system, Omnipage, in a pilot study with over 1000 expressions from images in 3 high school and college math books, for more accurate text and Braille output and therefore less human editing. With less editing needed, more translation can be done in the same amount of time, leading to quicker access to resources. This combined approach has the advantage of being at least as good as both recognition engines for each book so that the user does not have to worry about figuring out which recognition engine to use on a perbook or paper basis. Preliminary pilot study results show up to a 15% recognition improvement compared to InftyReader and Omnipage performing on their own, warranting further study on more textbooks.