all: add dpi parameter as manual override to image metadata

[bertsky]

Fixes #102.

This does influence segmentation and OSD a lot, but not so sure about (LSTM) recognition. Maybe we should document this somewhere – it's one of very few parameters we have to influence the quality of Tesseract's layout analysis (i.e. deliberately setting a fake value can improve results).

Note that Tesseract is optimised for modern fonts (which are typically smaller than historic ones) and thus biased against historic prints. So setting higher than factual DPI could be a useful recommendation for block (and maybe even line) segmentation. But this is still conjecture and I have only little evidence supporting it – someone would have to make systematic measurements first.

[codecov]

Codecov Report

Merging #108 into master will decrease coverage by 1.57%.
The diff coverage is 7.14%.

@@            Coverage Diff             @@
##           master     #108      +/-   ##
==========================================
- Coverage   39.14%   37.57%   -1.58%     
==========================================
  Files           9        9              
  Lines         894      942      +48     
  Branches      190      204      +14     
==========================================
+ Hits          350      354       +4     
- Misses        492      528      +36     
- Partials       52       60       +8

Impacted Files	Coverage Δ
ocrd_tesserocr/segment_table.py	0% <0%> (ø)	⬆️
ocrd_tesserocr/crop.py	12.93% <0%> (-0.84%)	⬇️
ocrd_tesserocr/deskew.py	16.19% <0%> (-1.16%)	⬇️
ocrd_tesserocr/segment_word.py	72.88% <12.5%> (-7.89%)	⬇️
ocrd_tesserocr/segment_line.py	70.76% <12.5%> (-6.82%)	⬇️
ocrd_tesserocr/recognize.py	50.72% <12.5%> (-0.96%)	⬇️
ocrd_tesserocr/segment_region.py	57.03% <12.5%> (-2.48%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8ccde94...3a684a5. Read the comment docs.

[bertsky]

@kba when you merge, please don't forget to update CHANGELOG.md this time.

As for versioning, I think 108, 109 and 110 are all new features, therefore I suggest 0.8.0.

[kba]

please don't forget to update CHANGELOG.md this time

I'll try to, but I'm grateful for contributions. You have the best overview to outline your PR in broad strokes. Updated for 0.7.0 in master.

[bertsky]

I'll try to, but I'm grateful for contributions. You have the best overview to outline your PR in broad strokes.

ok, first, for 0.7.0, I believe these need to move from Added to Changed:

• interprete overwrite_regions more consistently, improve segmentation #104
• annotate @orientation (independent of dedicated deskewing processor) for vertical and @type for all other text blocks, improve segmentation #104
• no separators and noise regions in reading order, improve segmentation #104

then for 0.8.0, how about this:

• Added:
  • parameter dpi overriding pixel density meta-data in all processors, Additional parameter for DPI override #102
  • parameters char_whitelist, char_blacklist, char_unblacklist in recognize, recognize: expose character white/blacklisting parameters #107
  • set lstm_choice_mode=2 when textequiv_level=glyph to get character alternatives from LSTM in recognize, also fill PAGE's glyphs and its variants and confidences via GetIterator() in recognize.py  #7