Improve FFR detection rate

[abraker95]

Look into training the OCR to detect the NotoSans font better. I am using tesseract version 5.0.0-alpha.20200328

See:

• https://github.com/tesseract-ocr/tessdoc/blob/master/TrainingTesseract-4.00.md
• https://www.endpoint.com/blog/2018/07/09/training-tesseract-models-from-scratch
• https://www.youtube.com/watch?v=TpD76k2HYms
• https://pretius.com/how-to-prepare-training-files-for-tesseract-ocr-and-improve-characters-recognition/
• https://github.com/tesseract-ocr/tesseract/issues/1174#issuecomment-338448972

[abraker95]

So from initial research:

• This needs to be done in linux as there is no support to train it on windows. Luckily windows has linux WSL (Windows Subsystem for Linux)
• From the .ttf font generate .tif images containing of a bunch of random characters. I already have that working, just need to implement saving it to file.
• .box files also go along with each .tif image. *.box files have a list of data defining bounding box for each character in the image.

Some nuances concerning box files:

• The (0,0) xy-coordinate is at bottom-left
• Each box file has to has to be named the same as its respective TIFF image counterpart
• The format of box file is: <symbol> <left> <bottom> <right> <top> <page>, where symbol is character, left bottom right and top are px coordinates. page should be 0 unless TIFF files are multi-page.

Sample box data:

d 112 4654 136 4690 0
i 141 4654 147 4690 0
f 149 4654 166 4691 0
f 163 4654 180 4691 0
e 179 4653 204 4681 0
r 208 4653 224 4680 0
e 224 4653 249 4680 0
n 253 4653 276 4680 0
t 279 4653 292 4688 0
  292 4653 310 4689 0
N 310 4653 339 4689 0
e 344 4653 369 4680 0
w 368 4652 408 4679 0