Kiuyha/Manga-Bubble-YOLO

Manga Text Bubble Detector (YOLO-Nano)

This repository contains a lightweight object detection model designed to detect speech bubbles and text regions in Manga pages. It is useing YOLO26 architecture that utilizes an End-to-End (Head-to-Head) prediction head, eliminating the need for Non-Maximum Suppression (NMS) during inference. It being trained on a diverse dataset of English, Vietnamese, and Japanese manga.

Dataset Details

The model was trained on a composite dataset containing 5,595 images, utilizing a mix of high-quality scans and fan translations. The data was split into 80% Train, 10% Validation, and 10% Test.

Folder Structure & Counts

/Manga_Project
├── images/
│   ├── train/ (4,416 images)
│   ├── val/   (579 images)
│   └── test/  (600 images)
└── labels/ (Corresponding YOLO .txt files)

Data Sources

• Manga109-s: ~3,000 images (High-quality official scans)
• Mangadex-EN: ~2,000 images (English fan translations)
• Mangadex-VI: ~1,000 images (Vietnamese fan translations)

Note: Some Mangadex images were filtered out during processing due to download errors or API limits, resulting in the final count of ~5.6k images.

Model Performance

The model was trained for 100 epochs (with early stopping) on an Nvidia Tesla T4 at 1280x1280 resolution.

Metrics Comparison (Test Set - 600 Images)

Model	Class	Precision	Recall	mAP@50	mAP@50-95	Params
YOLO26n	Text	0.929	0.863	0.947	0.765	2.4M
YOLO26s	Text	0.937	0.893	0.961	0.802	9.5M

• Inference Speed (T4 GPU):
  • Nano: ~11.0ms per image
  • Small: ~27.5ms per image

Usage

1. Using PyTorch (Python)

Requires ultralytics package.

from ultralytics import YOLO

# Load the model
model = YOLO('model.pt')

# Run inference
# Note: imgsz=1280 is recommended for small text bubbles
results = model.predict('path/to/manga_page.jpg', imgsz=1280, conf=0.25)

# Display result
results[0].show()

2. Using ONNX (Python)

Useful for deployment without PyTorch dependencies.

import onnxruntime as ort
import numpy as np
import cv2

# Load model
session = ort.InferenceSession('model.onnx')

# Preprocess Image
img = cv2.imread('test.jpg')
img = cv2.resize(img, (1280, 1280))
img = img.transpose((2, 0, 1)) # HWC -> CHW
img = np.expand_dims(img, axis=0) # Add batch dimension
img = img.astype(np.float32) / 255.0 # Normalize

# Run Inference
input_name = session.get_inputs()[0].name
outputs = session.run(None, {input_name: img})

print("Output Shape:", outputs[0].shape)
# Returns (1, 300, 6)

Example Result in YOLO26-n

Test 1

Test 2

Test 3

Training Configuration

The model was trained using the following hyperparameters:

model.train(
  data='dataset/data.yaml',
  epochs=100,
  patience=10,
  batch=8,
  lr0=0.0001,
  imgsz=1280,
  device='cuda'
)

Credits & Citations

We gratefully acknowledge the following datasets and tools used to build this project:

Manga109-s Dataset Used for high-quality Japanese manga panel data.

@article{multimedia_aizawa_2020,
  author={Kiyoharu Aizawa and Azuma Fujimoto and Atsushi Otsubo and Toru Ogawa and Yusuke Matsui and Koki Tsubota and Hikaru Ikuta},
  title={Building a Manga Dataset ``Manga109'' with Annotations for Multimedia Applications},
  journal={IEEE MultiMedia},
  volume={27},
  number={2},
  pages={8--18},
  doi={10.1109/mmul.2020.2987895},
  year={2020}
}

Magi (Annotation Tool) Used to auto-annotate the Mangadex portion of the dataset.

@misc{magiv2,
  title={Tails Tell Tales: Chapter-Wide Manga Transcriptions with Character Names}, 
  author={Ragav Sachdeva and Gyungin Shin and Andrew Zisserman},
  year={2024},
  eprint={2408.00298},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={[https://arxiv.org/abs/2408.00298](https://arxiv.org/abs/2408.00298)}, 
}

Note: The dataset cannot be released due to copyright concerns.