Is there a good way to change the recognition rate for such images?

[Cain Pian]

When PSM=6, close characters are concatenated.

When PSM=11, single digits are not recognized.

When PSM=12, single digits are recognized as vertical text.

I have trained thousands of similar images but this problem has not improved, is there a suitable parameter or method to solve this problem?

[Jeremiah]

I do not believe training would have any impact on whether or not the column layout is correctly identified during the page segmentation step.  I have similarly experienced the issue with single-digit columns being misidentified as vertical text when running with PSMs that use automatic page segmentation, so can confirm this is a systemic issue and not just something weird about this specific input.  Unfortunately, I am not aware of an existing option that prevents Tesseract from recognizing vertical text during automatic page segmentation, so this would probably require an additional option.  Would probably not be that hard to implement. 

[Cain Pian]

Yes, I've seen a lot of discussion on this issue that ended up going nowhere, it might be helpful to know what part of the code is affecting this.