This week, I was able to essentially finalize my work on the Tesseract for use with OCR. I have utilized hundreds of public, research grade documents and images, including ~200 from FUNSD (Form Understanding in Noisy Scanned Documents) and ~1.3k from Tobacco800 (complex document image processing test set from Illinois Institute of Technology), which seemed to have about a 95% accuracy rate. I suspect this is largely because stock Tesseract is trained mainly on printed text, so its models and language priors don’t match handwritten cursive, which is what is often in FUNSD and Tobacco800. When I tested it on the Open Food Facts dataset, which is much closer to our use case (printed, commercial packaging to be scanned) and closer to what Tesseract was trained on, the accuracy rate was closer to 99%, which is our intended goal.
I’ve also written the bulk of the code for mapping Braille dot patterns to actuator sequence, but still need to more comprehensively test this, using both written test cases and with the physical BrailleMate device.
My progress is on schedule, and I am excited for how the project has come together! In my group, I have been responsible for much of the software, which has been interesting to learn about. In coming weeks, I will need to keep testing Tesseract on physical bottles, not just online datasets. Now that Abby has largely set up the Raspberry Pi, I will also need to update the Braille dot pattern code to work with our hardware, rather than just my Python test cases.