Made a japanese ocr for manga

the model is only 230mb , but the accuracy is pretty good

it is still in development and it’s in alpha stage, i’m rewriting it in c++ because the framework ncnn is written in that language, and because one of the goals was to have a fully c++ ocr.

image OCR result
当民宿自慢の春の海鮮料理づくし!

the technology used is pretty recent, the format is “pnnx” which was developed by Tencent Inc.
The only ocr that can beat mote-ocr as by now is manga-ocr, but it relies on a bunch of obscure libraries and the ecosystem is fragile (one of the reasons why i prefer not having external dependencies)

advantages of mote-ocr over manga-ocr:

  • manga-ocr needs 7gb of space, while mote-ocr needs only 230mb
  • no need to install third-party libraries, no headaches with updates.
  • self-contained codebase, you know what you’re running
  • optimized for low-end computers or even embed devices (aka can run in a potato)

Thanks to the chinese community who maintains ncnn and Thanks specially to Anya, who helped me figure out how the model worked.

3 个赞

Let’s put this topic in Tech Share category.