Hi si dinh, thank you for read

I uploaded the results of your images to the repository in github and I had two parameter changes, since the images are small and don’t have much shade.
The first one, I resized the image to 1024 in width (I only removed one condition for this)
The second, I created a new parameter to be able to disable the light compensation function (binarization step).

To improve line segmentation, you can change the parameters of chunksNumber and chunksProcess too, as required by your application.

All these parameters are in the file (unfortunately I didn’t put easy access via python): src /imgproc/cpp/main.cpp

The words are already saved automatically within the output folder of each image.

And I saw that you highlighted the part of the text on the deslant method. The improvement was based on the cursive text recognition task. With the tests I did here, I realized that it also improves for this type of line segmentation (which uses histogram projection on the y-axis). But for those images that you have passed (printed texts in general), I believe that doesn’t change the result in the end.

Best.