manga-ocr-android

Manga OCR made fast

What?

This repository is an attempt to port the work from manga-ocr and koharu to Android, for integration into Mihon. See the model-dev README for more details on the model work.

How?

The library module is published via jitpack. See badge above for latest version and extra install instructions.

dependencies {
    implementation 'com.github.dhleong:manga-ocr-android:<latest-version>'
}

The required models are downloaded automatically when first used.

Is it really fast?

I haven't done any real benchmarking, but anecdotally—yes! A quantized version of the comictextdetector model from koharu took about 3s on average to detect text boxes; the default Detector implementation in the library works in about 300 milliseconds.

A quantized version of the manga-ocr model took about 110ms per character in the box. The current MangaOcr implementation in the library only takes about 10ms per character.