ISSUE: Loss of black text clarity in white-dominant images when scaling 2x, 3x, 4x

[fusefib]

Description: I am experiencing difficulties with the normal binarizing feature in ScanTailor-Experimental. When attempting to process an image containing predominantly white background with little black text, the resulting binarized image displays faint or erased text when scaling up the resolution to 2x or higher. Adjusting the Curve/Sqr color settings from 0.5/0 to 1/1 provides some improvement but does not fully restore the text clarity. Changing the Threshold method from Otsu does, but keeping Otsu is preferred. Setting the thickness to the max doesn't help.

Expected Behavior: The binarization process should maintain the clarity and legibility of the black text, even when scaling up the resolution.

Actual Behavior: The black text becomes faint or disappears entirely in the binarized image when the resolution is increased.

Environment:

• ScanTailor-Experimental version: 0.2024.05.03-X86-64-Qt6 ("threshold" from Releases)
• Operating System: Windows 10

Steps to reproduce:

• Add attached image to a new project.
• Go directly to Stage 6 - Output.
• Check Resolution Enhancement box and click 2x.
• 4x disappears most of the black.

Notes: I've just started using this variant of ScanTailor (awesome work!) and I haven't done extensive testing to determine if this is a regression and, if so, when it occurred. Would be happy to test if needed.

[zvezdochiot]

Hi, @fusefib .

In Linux, the one you described cannot be played either on 2x or 4x.

[fusefib]

I'll try to prepare a better picture for testing. In any case, it consistently does this for high resolution pictures that are mostly white and have almost no black, when scaling up. However, pages with enough text always turn out fine. Here is the result for me: testpic-scaled.zip

2x: 4x:

[plzombie]

@zvezdochiot Reproduced with Otsu, Dots8, EdgePlus(somehow), BlurDiv, EdgeDiv, Robust Not reproducible in 0.2023.10.24 Reproducible in 0.2023.10.30 and so on Looks like some bug introduced in 0.2023.10.30 (for all sizes) and partially fixed in 0.2023.11.19 (for size 1.0). Also if you change image size manually by 50% and then upscale 2x in ScanTailor - everything is ok, but when you upscale 4x you will got the same result like original image with scale 2x (thin black lines). So its probably something related to image size.

Also reproducible on 32-bit linux. So maybe should look for long (32bit on windows and 32bit linux, 64bit on 64bit linux) variables

I suggest to check https://github.com/ImageProcessing-ElectronicPublications/scantailor-experimental/commit/aafb67f95a0685535c999ede9fb5f8ed5112beed https://github.com/ImageProcessing-ElectronicPublications/scantailor-experimental/commit/587683bdda23650fa0d237df6ebe85db434f2a13 and https://github.com/ImageProcessing-ElectronicPublications/scantailor-experimental/compare/0.2023.11.10...0.2023.11.19#diff-4ddd4b67d244ac7c7864d278c0179f232b36742520daee44b20ec5b80e4f6324R870 (and maybe other code) and replace long with int64_t/uint64_t

[zvezdochiot]

@plzombie , this could be (?!) an overflow in: https://github.com/ImageProcessing-ElectronicPublications/scantailor-experimental/blob/11d943af4b93c44e6b89bf47f90388025009e30f/src/imageproc/Binarize.cpp#L197 but the type is specified independent of the bit depth (!).

PS: All the thresholds you listed use binarizeBiModalValue.

1. :question: Maybe: https://github.com/ImageProcessing-ElectronicPublications/scantailor-experimental/blob/11d943af4b93c44e6b89bf47f90388025009e30f/src/imageproc/Binarize.cpp#L244-L245 to

       Tb = (ib > 0) ? (Tb / (double) ib) : 0.0;
       Tw = (iw > 0) ? (Tw / (double) iw) : (double) histsize;

2. :question: Maybe: https://github.com/ImageProcessing-ElectronicPublications/scantailor-experimental/blob/11d943af4b93c44e6b89bf47f90388025009e30f/src/imageproc/Binarize.cpp#L230-L243 to:

       for (k = 0; k < histsize; k++)
       {
           im = histogram[k];
           if (k < threshold)
           {
               Tb += ((double) im * (double) k);
               ib += im;
           }
           else
           {
               Tw += ((double) im * (double) k);
               iw += im;
           }
       }

3. :question: Maybe: https://github.com/ImageProcessing-ElectronicPublications/scantailor-experimental/blob/11d943af4b93c44e6b89bf47f90388025009e30f/src/imageproc/Binarize.cpp#L197 to:

   unsigned long int im, histogram[histsize] = {0};
   double iw = 0.0, ib = 0.0;

[plzombie]

@fusefib Could you try this build? staging.zip scantailor-experimental-0.2024.05.05-X86-64-Qt6-install.zip

[fusefib]

Confirmed it works properly now with both zips. Thanks to both of you for the fix!