[
https://issues.apache.org/jira/browse/TIKA-4947?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18125198#comment-18125198
]
Tim Allison commented on TIKA-4947:
-----------------------------------
I looked around a bit and couldn't find anything that didn't cause more damage
than it helped.
I was particularly hopeful that PP-LCNet would work. But it was more wrong more
than right on rendered images from the first pages of PDFs from digital
corpora. I looked at 2k pages. There were particular weaknesses with RTL
language pages.
tesseract osd with only orientation detection is wrong more than it is right.
BUT, when it is run with OCR, it is quite accurate, at least on the files I ran
against. So, we can't run only orientation detection with tesseract and be done.
At this point, I think we have to rely on the vlm model to handle this. This is
not something that we can reliably fix before sending images off to models.
This answer does not bring me joy.
> Assess how we're doing with orientation detection for OCR
> ---------------------------------------------------------
>
> Key: TIKA-4947
> URL: https://issues.apache.org/jira/browse/TIKA-4947
> Project: Tika
> Issue Type: Task
> Reporter: Tim Allison
> Priority: Major
>
> Some vlm-based OCR tools are not robust against rotation and can hallucinate
> completely fluent content if an image is rotated.
> As with the blank image detection (TIKA-4946), I'm not sure if we put this in
> the PDFParser or the OCR engine or both.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
