OCR performance

[michel_2440]

Hi

I have been using Acrobat since years either for reading or for getting editable text from scans sometimes 100 pages documents. 

I experience too many errors in OCR so that when i copy/paste recognized text into external applications (usually Microsoft Office) i have to spend a significant amount of time fixing the text. That even when the scan has a pretty good quality and low level of noise or distorsion.

Recently i loaded pdf files (whose text was previously recognized by Acrobat OCR) into multiple AI's (chatGPT, NotebookLM, Mistral, etc). It appears that these AI's where doing à better recognition in locations where Acrobat text was wrong. 

So my question is whether build-in Acrobat OCR performance is going to improve and provide much better end user experience. This is indeed a serious matter of productivity since i would not like to be forced to use external tools to workaround errors in Acrobat OCR.

Thanks for your feedback

[michel_2440]

I complete my point by saying OCR is a quite basic process while context is key to solve either syntactical and semantic ambiguities. So i would expect built-in capability to reach a récognition level without spending hours fixing errors

[Randy Hagan]

I would contend, from lots of experience, that OCR is far from a basic process, with lots of variables within the software to consider even before considering the variables of what you're scanning and how compatible it will be with your OCR setup. 

There's no substitute for proofing your OCR work when you use it, even if you're pouring it into AI tools to reduce the proofing time. Certainly AI is good at making assumptions at filling the gaps, but you're trading the obvious errors you can see and correct from your OCR for possible mistaken assumptions that are technically accurate but susceptible to potential autocorrect errors that will read through spellchecking just fine.

The folks here are end users, just like you. This writer included. So we don't have any unique input to Adobe or sway over how the company develops its products. If you'd like to submit your concerns to Adobe developers, you can do that with Adobe UserVoice through this link.

But I can tell you, there are no shortcuts to doing a thorough proofing of your OCR output. I've tried your workflow, and I find it's a lot easier for me to find and fix the obvious mistakes that occur with OCR output than it is to first run it through AI tools and have to find misinterpretations/"hallucinations" that read through a cursory proofing but are called out when doing a thorough proof on my copy.

Hope this helps,

Randy

[gary_sc]

As someone who's been working with OCR for about 30 years, it's always been amazing to me it's as accurate as it is. 

That notwithstanding, I agree that it could and should be better. I'm hoping that with AI, the quality and accuracy will improve. (I'm still waiting)

Meanwhile, what you get out of OCR currently is significantly dependant upon what you put into it. I encourage you to read through this blog I wrote for Adobe a number of years ago, it still holds up today.

https://community.adobe.com/t5/adobe-community-professionals/scanning-clean-searchable-pdfs/m-p/4785...

Good luck!