Characters swapped in results from 3.05.01

[QuLogic]

With 3.05.01, the basic English test from pyocr started failing. It supports both OCR via the library or calling out to tesseract. The basic test uses this file, does some conversion, and passes it along to tesseract. I have attached the bitmap file that goes directly into tesseract, though it also works if you pass the png unchanged. The expected text is here. The effective command is: tesseract input.bmp output -l eng -psm 3.

Environment

• Tesseract Version: 3.05.00 / 3.05.01
• Commit Number: 8ff183b / acf318a
• Platform: Fedora 27 x86_64

Current Behavior:

With 3.05.01, the text comes out as expected except the word "ocr" comes our as "cor".

Expected Behavior:

With 3.05.00, the text matches entirely and tesseract produces "ocr".

Suggested Fix:

I have bisected this change to be somewhere between 8ff183b and acf318a. I could not bisect any further because the commits did not compile (bisect is a great reason to ensure that commits compile!). As all the intervening commits appear to be code cleanup, I believe this change was an unintended regression.

[QuLogic]

If I go back to 8ff183b (the last working commit), then cherry-pick 90db5a1, it starts to fail again.

So, heading back to the earliest broken commit (or even the 3.05 branch), then reverting 90db5a1 fixes the problem. I don't really know the code so I can't say that's the exact fix, but it seems to work for this case.

[amitdo]

Thanks for the report and analysis.

90db5a1 was backported from 9a5ed19 (4.00). Ray Smith reverted it in da03e4e.

[stweil]

That sounds as if it has to be reverted for the 3.05 branch, too. @rfschtkt, what do you think?

[amitdo]

Let's talk next year :-)

[amitdo]

#529 (comment)

From a quick look, Dict::Load() doesn't add load_bigram_dawg to dawgs_, which either is a bug or should be documented in a comment

Ray also added this comment to Dict::Load()

  if (load_bigram_dawg) {
    bigram_dawg_ = dawg_cache_->GetSquishedDawg(lang, TESSDATA_BIGRAM_DAWG,
                                                dawg_debug_level, data_file);
    // The bigram_dawg_ is NOT used like the other dawgs! DO NOT add to the
    // dawgs_!!
}

[Shreeshrii]

@zdenop Please label
accuracy.

90db5a1 was backported from 9a5ed19 (4.00). Ray Smith reverted it in da03e4e.

That sounds as if it has to be reverted for the 3.05 branch, too.

[QuLogic]

Soo, is there any chance the fix will be backported?

[zdenop]

is it needed? There is no plan to release new 3.05. We are focused on 4.0 release.

[amitdo]

This issue should be fixed or closed.

[zdenop]

Can you please check 3.05 branch?

[QuLogic]

Yep, looks like it's working there now.