Model Description

This is a fine-tuned version of the PaddleOCR v5 Server Detection Model. It has been trained on a dataset of manga speech bubble crops to improve detection for:

• Speech Bubbles Lines: Standard dialogue detection.
• Vertical Text Lines: Improved bounding boxes for Japanese vertical writing (tategaki).
• Text Lines Outside Bubbles: Narration boxes and floating text.
• Text Lines With Furigana: Greatly reduced the creation of separate bounding regions for furigana.

This model outputs bounding boxes (polygons) for text regions. It does not perform text recognition; you will need a separate recognition model for that.

Note that this model is still being worked on, and may improve with a better dataset or hyperparameters.

Training Data

The dataset consisted largely of synthetic data due to the limited real samples available.

• ~400 randomly sampled speech bubble crops from Manga109s
• ~200k synthetic images

Acknowledgments

This project was done with the usage of:

• Manga109-s dataset
• CC-100 dataset
• MangaOCR synthetic data generation (code was edited for speedups, bounding box additions, and improved representation of manga)