@egorpugin, please wait a little bit longer before merging non-trivial pull requests. Otherwise there is not enough time for others to review the changes.

I am not sure that this change is the right solution. punct_stripped always could produce intervals where end was smaller than start. Dict::valid_bigram has special handling for start >= end. Now start > end no longer can happen, but those cases now return start == end, so the tests still work.

Therefore the fix is okay.

@EnodoGH, can you provide a test image which triggers the crash?

~ punct_stripped always could produce intervals where end was smaller than start. Dict::valid_bigram has special handling
for start >= end. Now start > end no longer can happen, but those cases now return start == end, so the tests still work.

Therefore the fix is okay.

Yes, I think this is also the right place to do it, because otherwise you continue having different kinds of empty intervals for no good reason. Neither EqualIgnoringCaseAndTerminalPunct() nor Dict::valid_bigram() needs them semantically, which is why it feels a bit artificial to change EqualIgnoringCaseAndTerminalPunct() to expect them.

@EnodoGH, can you provide a test image which triggers the crash?

I probably could, but it would be highly non-trivial. It's one of slightly fewer than one million images, which our software takes a few days to process. There's a good chance it's a customer image, for which I'll need to get permission before passing it on to you. In that image, there are multiple calls to the Tesseract library, and I'd have to isolate the exact call and the exact data that was passed to it.

I'll try to do all that if it's too hard for you to reproduce the crash from scratch. For now, all I can tell you is that we reproduced the crash in EqualIgnoringCaseAndTerminalPunct() on two different machines. On mine, I managed to identify punct_stripped() as the reason the crash happened, fixed it, and the crash was gone for good on both machines.