PP-OCRv6 small rec — manga fine-tune
PaddlePaddle's PP-OCRv6 small text recogniser, fine-tuned on a manga corpus for Japanese manga lettering — vertical text, small kana, elongation marks, decorative symbols, stylised fonts — then interpolated back toward the base weights so it keeps its general Japanese OCR ability. Same architecture, input contract, dictionary and CTC decode as the stock model: a 21 MB drop-in replacement.
Results
Scored per bubble on 422 hand-transcribed manga speech bubbles, held out from training:
| stock PP-OCRv6 small rec | this model | |
|---|---|---|
| Mean CER (strict) | 0.0615 | 0.0402 |
| Mean CER (lenient) | 0.0449 | 0.0343 |
| Bubbles read exactly | 63.3 % | 75.1 % |
Downstream, the better transcriptions were worth +3.4 chrF++ through a Japanese→English manga translation model. Against larger whole-bubble OCR models (up to 30× the parameters), this 5.3 M-parameter model stays within a few thousandths of CER at a fraction of the compute.
Usage
- Input
[N, 3, 48, W], RGB, height 48 with aspect preserved, normalised(x / 255 − 0.5) / 0.5. - Output
[N, T, 18710]: stockppocrv6_dict.txt+ space + CTC blank (index 0); greedy CTC decode. - Recogniser only — pair it with a PP-OCR detector and feed it line crops.
- ONNX opset 13; Softmax axes set to −1 so it runs under ONNX Runtime's XNNPACK provider on Android.
SHA-256: c5cc5038a98c3df3e2d37de5716f603e2b0bcd3536c74078fdd91876a48a25ef
Limitations
Japanese only. Furigana is not suppressed by the model; drop ruby lines before recognition. Rare kanji outside the manga distribution and the ♥ glyph remain the most common residual errors.
License
Apache-2.0, as the base model.