Tesseract 5.0 OCR Engine Bringing Faster Performance With "Fast Floats"

[phoronix]

Phoronix: Tesseract 5.0 OCR Engine Bringing Faster Performance With "Fast Floats"

Tesseract as the leading open-source optical character recognition (OCR) engine that employs neural networks for converting images/scans of text into actual recognized text is nearing its 5.0 release...

https://www.phoronix.com/scan.php?page=news_item&px=Tesseract-5.0-Beta

[aufkrawall]

Curious about AVX512 optimizations, it makes a dramatic difference in libdav1d with 10 bit content (as Rocket Lake kills Zen 3 here). If this will be true for Tesseract too, its usefulness probably should be re-evaluated.

[elatllat]

LOL, I'm finishing a week long OCR run now. It's not obvious if fast float is enabled in the beta binaries or how to enable it in the build but maybe it's worth looking into...

[coder]

Traditionally the Tesseract OCR engine has relied upon doubles

WTF??

Did these clowns actually write their own deep learning framework? If so, they should ditch it and use one of the others, for probably an overnight order-of-magnitude speedup. Not exaggerating.

[coder]

Curious about AVX512 optimizations, it makes a dramatic difference in libdav1d with 10 bit content (as Rocket Lake kills Zen 3 here).

Look at other CPU-based deep learning benchmarks. OpenVINO is a good one, because both its AVX2 and AVX-512 paths are well-optimized. Michael has some other deep learning benchmarks in PTS that are pretty rubbish, in that they don't represent realistic performance on CPU.

As for DAV1D, that's another rubbish comparison, simply because the non- AVX-512 path for 10-bit wasn't comparably optimized. So, it's not an apples-to-apples comparison.

[aufkrawall]

More optimization for non-AVX512 path in libdav1d is only a matter of time then? It's been a while since they started using it.

[Joe2021]

So why were double instead of float used in the first place? Out of a habit?

[Chewi]

WTF??

Did these clowns actually write their own deep learning framework? If so, they should ditch it and use one of the others, for probably an overnight order-of-magnitude speedup. Not exaggerating.

I don't know whether theirs is any good or not, but you should know that Tesseract was started in the 1980s, making it waaay older than any of the other things you're thinking of.

[Aryma]

Tesseract always give me funny result when i OCR some japanese document

[coder]

More optimization for non-AVX512 path in libdav1d is only a matter of time then? It's been a while since they started using it.

Someone probably has to fund the work. I forget who, but someone paid them to do the AVX-512 optimization of the 10-bit path. I think it was like Netflix or someone like that -- not Intel.