Automated tactile graphics translation: in the field
Chandrika Jayant, Matt Renzelmann, Dana Wen, Satria Krisnandi, Richard Ladner, Dan Comden · 2007 · Proceedings of the 9th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '07) · doi:10.1145/1296843.1296858
Summary
This paper describes two years of field work extending the Tactile Graphics Assistant (TGA), a system developed at the University of Washington to automate the translation of figures from mathematics, science, and engineering (STEM) textbooks into embossed tactile graphics for blind students. Building on the authors' 2005 paper, which introduced the TGA concept, this work reports on the refined production workflow, new software features, and real-world translation of more than 2,300 figures across four textbooks (computer architecture, precalculus, astrophysics, and discrete mathematics). The workflow combines scanning, image preprocessing in Photoshop, text extraction via the TGA, optical character recognition (OCR), Braille translation using Duxbury or Braille2000, and final assembly in Adobe Illustrator, with batch processing used wherever possible to accelerate throughput. The authors explain why standard OCR alone cannot solve the problem (it misidentifies graphical elements as text and cannot reinsert Braille at the correct locations), and describe how the TGA uses support vector machine classification on radial-density features of connected components to reliably distinguish characters from graphics. New features described include automated label placement via a priority-queue optimization algorithm, text rotation to horizontalize angled labels for OCR, and integration with InftyReader for mathematical content. The paper also documents the authors' April 2007 training workshop at the National Braille Association conference, where 60 tactile graphics specialists and transcribers were introduced to the tools.
Key findings
Across three books with detailed timing data, figures were translated in an average of 6.3 to 10.2 minutes of human time per figure, a dramatic reduction from traditional manual methods that take hours per figure. The character-finding algorithm achieved a 0.68% error rate (92 false positives and 17 false negatives across 16,145 connected components), and the label training algorithm misgrouped only 13 of 824 labels on test figures. Label placement emerged as the most time-consuming step, consuming 23.3% of total translation time on average, which motivated development of the new automated placement algorithm. Illustrator-based assembly and OCR editing were the other dominant time sinks, particularly for math-heavy books where regular OCR performs poorly. The authors found that training the TGA required only a few representative images per batch; in one case, under five training images successfully handled a batch of 600 figures. Setup costs (scanning and cropping) were disproportionately high for print-sourced books (up to 18.3% of time) compared to books available in digital form. Training workshop participants responded positively, and at least 20 practitioners downloaded the TGA software following the workshop.
Relevance
This paper is a landmark in practical STEM accessibility: it demonstrates that automated tactile graphics production can move from lab prototype to field-deployable tooling for working transcribers. For accessibility practitioners, the key lesson is that general-purpose tools (commercial OCR, generic image processing) are insufficient for domain-specific accessibility tasks; domain knowledge about textbook figure conventions enables far better automation. The workflow decomposition and timing data remain instructive for anyone planning alternative-format production pipelines. Limitations include dependence on a specific embosser (Tiger Embosser), substantial manual editing still required at each step, and weak support for complex mathematics without additional math-OCR tooling. The need to re-train models per textbook class also limits generalization. Nearly two decades on, the underlying challenges (scanned-image noise, math recognition, visual-to-tactual simplification) remain relevant to modern tactile graphics work and to emerging AI-assisted alternative-format pipelines.
Tags: tactile graphics · Braille · STEM accessibility · blindness and low vision · image processing · machine learning · OCR · automated alternative formats · mathematical accessibility · accessible publishing
Standards referenced: Nemeth Braille · Grade 2 Braille