Regarding Tesseract Alternative

[Devansh Varshney]

Hi,

I am a contributor to LibreOffice and I have as of now used teserract based extension https://extensions.libreoffice.org/en/extensions/show/99360

and here is the GitHub repo - https://www.github.com/varshneydevansh/TejOCR

The thing what I have noticed is that tesseract is not that reliable when it comes to the maintain the structure of the OCR atleast and even felt that the accuracy is bit off.

I am looking for some alternatives given the advancements in Image Processing nowadays.

It could be either based on DocTr. https://www.github.com/mindee/doctr

or could be a VLM based derivative solution - https://www.huggingface.co/Qwen/Qwen2-VL-2B-Instruct
https://www.huggingface.co/blog/smolvlm#:~:text=such%20as%20a%20laptop!%20You,which%20use%20a%20similar%20approach

https://www.huggingface.co/microsoft/Florence-2-large#:~:text=OCR%20with%20Region

https://www.huggingface.co/AI-Safeguard/Ivy-VL-llava

What I am looking or is what could be the best way which uses as less compute as possible and can run entierly locally atleast by CPU.

Regards,
Devansh