Library support for training tools

[stefan6419846]

Environment

• Tesseract Version: 4.1.3, but affects latest main branch as well
• Platform: Linux localhost.localdomain 5.3.18-150300.59.68-default #1 SMP Wed May 4 11:29:09 UTC 2022 (ea30951) x86_64 x86_64 x86_64 GNU/Linux

Current Behavior

Using the training tools from src/training as a library/shared object is impossible or requires lots of manual modifications for each upstream change in this repository.

Current limitations include (but are not limited to):

• Duplicate names (main() and parameter flags).
• Parameters have to be passed as argc and argv in some cases where parameter handling is done without flags, while flag-based parameters are not nice to handle either.
• In some source files exit(1) is used inside main(), instead of return 1 which generally should be equivalent.

Expected Behavior

The training tools can be used as a library/shared object with a well-defined API like it is possible with the regular Tesseract code. This allows developers to use the training tools in their own projects, for example by wrapping the native C++ code into a language-specific package (like a Python module) to avoid subprocess calls.

Suggested Fix

• Introduce a header for the training functionality.
• Make implementations independent from main() and allow calling them with regular parameters.

[stweil]

The current build already creates libtesseract_training.a. Is that library sufficient (maybe if in addition a shared library libtesseract_training.so is built)?

[stefan6419846]

Depending on the use case, either a static or a shared library may make sense, so supporting both (if feasible) probably is the way to go in my opinion.

As for the original request, a suitable header file seems to be still missing. Additionally, the issues regarding duplicate names and the non-optimal behavior for providing parameters appear to still be unresolved.