update leptonica, tesseract, and libk2pdfopt

[benoit-pierre]

Depend on https://github.com/koreader/libk2pdfopt/pull/49.

Thanks to some upstream cleanups in tesseract, the resulting code size (bss+data+text) is reduced by ~1.4-2.3 MB.

Note: tesseract is compiled with legacy engine support, which means the resulting mode after initialization will depend on the type of language data available:

• tessdata (the one we use for the testsuite): 22.4 MB for English; contains data for both legacy & new engine: run the LSTM recognizer (new), but allow fallback to Tesseract (legacy) when things get difficult.
• tessdata-fast: 3.92 MB for English; best “value for money” in speed vs accuracy (Integer models): run the LSTM line recognizer only.
• tessdata-best: 14.7 MB for English; best results on Google’s eval data, slower (Float models): same mode as above.

AFAIK, there are no legacy only language files.

I tested both modes, English only with tessdata & tessdata-fast, and it did not seem to make a difference (which might just mean that the fallback on legacy was never triggered). Disabling the legacy engine would save an additional ~500 KB. But I don't know if the new engine is (always) better, especially for other languages.

The version used by Linux distributions seem to vary: for example Arch Linux is using the English tessdata variant, while Ubuntu uses tessdata-fast. ¯\(ツ)/¯

---

This change is 

[Frenzie]

I tested both modes, English only with tessdata & tessdata-fast, and it did not seem to make a difference (which might just mean that the fallback on legacy was never triggered).

You mean on a device I presume?

[benoit-pierre]

I tested both modes, English only with tessdata & tessdata-fast, and it did not seem to make a difference (which might just mean that the fallback on legacy was never triggered).

You mean on a device I presume?

Mainly with the emulator.