Ask HN: Why is OpenAI's OCR better than everything else? | Hacker News Reader

Ask HN: Why is OpenAI's OCR better than everything else?

3 pointslilouartz3 comments

This is a curiosity post/question.

I am trying to extract supplement nutrition data from photos of products. Over the last 24 hours, I've tried every popular OCR software and API, but none of them come even close OpenAI. With OpenAI Vision I was able to extract tabular data even from very obscure/low-res pictures. Meanwhile, OCR APIs offered by Microsoft, Amazon, Google, etc. and OpenSource softwares like tesseract, EasyOCR, etc. can barely extract the same information from screenshots.

What is so fundamentally different about how OpenAI extracts/interprets texts from the traditional OCR softwares?

Because it understands what it's looking at. No clearer way to put it.

If you compare scanning random garbage text, it will likely be as bad as the others, including humans.

I would like to hear from others how true this is.

My understanding (based on the wording that OpenAI itself uses) is that it still uses OCR tech behind the scenes. It seems like they have developed some models to better comprehend the same OCR scanned content that others produce.

It may well use OCR for extracting text, before cleaning it up.

Apple's "live text" OCR gave, on 2 lines of a curved, tilted, low-res image:

  Sert. size 1 Tosp (209)
  Total Sugars 12g
  Id. 9 Aded Supars 195%

When shown this, ChatGPT4 said "... This is where the context-sensitive interpretation I offer can be more effective, as I can understand that "Tosp" likely means "Tbsp" (tablespoon) and "Id. 9 Aded Supars" refers to "Incl. 9g Added Sugars."

Reply on news.ycombinator.com