PP-OCRv6 Medium Spanish Manga Recognition V2
This repository contains a custom ONNX export of a PP-OCRv6 medium text recognition model fine-tuned for Spanish manga OCR.
It is a recognition model only. Use it with a text detector such as PP-OCRv5_mobile_det, PP-OCRv6_medium_det, or another PaddleOCR-compatible detector.
Files
inference.onnx: ONNX recognizer model.inference.yml: PaddleOCR inference configuration and character dictionary.ppocrv6_medium_v2_spanish_manga_rec.tar: uncompressed browser asset archive containing inference.onnx and inference.yml.onnx_validation.json: ONNX Runtime validation summary.
Input/Output
- Input tensor:
x
- Input shape:
[batch, 3, 48, dynamic_width]
- Input type:
float32
- Output shape:
[batch, time, 18711]
- ONNX opset: 17
Notes
The model was trained for personal/research Spanish manga OCR experiments. It is strongest when paired with a page pipeline that uses CLAHE preprocessing and filters vertical/non-dialogue/SFX-like detections before or after recognition.
For browser use with PaddleOCR.js-style custom assets, use ppocrv6_medium_v2_spanish_manga_rec.tar.
Limitations
- This is not a detector.
- It may attempt to romanize Japanese SFX or non-Spanish decorative text if such crops are passed to the recognizer.
- Accents are improved but still imperfect; downstream Spanish normalization may help.
