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
