• PenisWenisGenius@lemmynsfw.com
    link
    fedilink
    English
    arrow-up
    3
    ·
    edit-2
    5 months ago

    There is a command line program called tesseract that does image to text generation. It produces plaintext from a picture of text. I didn’t look into exactly how it works but iirc, image to text that’s actually good and accurate needs ai shenanigans.

    • antler@feddit.rocks
      link
      fedilink
      English
      arrow-up
      2
      ·
      5 months ago

      It’s built by Google, but it’s open source, and is probably the best optical character recognition by far. It’s one pip/pipx installation away and I find it pretty useful on occasion. Same as WhisperAI by by OpenAI. Fully open source and one pip/pipx command away, probably close to the best audio transcription there is as well.

      Not sure either count as AI, at least not AI chatbot kind of AI more like more simple algorithms, but they’re great in the sense it’s just another program but a very useful tool. Not some baked in copilot kind of deal

      • iopq@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        ·
        5 months ago

        The algorithm is exactly the same as the chat bot, only the underlying data is different. Yes, they are all deep neural networks

      • evranch@lemmy.ca
        link
        fedilink
        English
        arrow-up
        2
        ·
        5 months ago

        probably the best optical character recognition by far

        I’ve actually just been working with OCR this week, trying to capture data off of the screen of a stupid proprietary Schneider device as that’s the only way to get at it.

        Long story short Tesseract stinks at this task.

        The Chinese designed PaddleOCR seems significantly superior as it runs a more modern neural net and requires a lot less preprocessing. I would class it as more of a “full service AI” and not just a simple recognition system like Tesseract, it can correct for skew and do its own normalization and thresholding internally while Tesseract wants a perfect boolean raster fed to it.

        Unfortunately, the barrier to entry is a lot higher due to trying to understand their text vomit website and the fact that it seems prone to random segfaulting.