[ML] Manga Segment

📃 | Description

This is a simple project developed in Python, aimed at removing manga backgrounds. I created this because I usually read manga at night.

To use this project in production, you need to install the Bandwidth Hero browser extension. If you prefer reading manga in a specific reader, I recommend TachiyomiAZ, which is compatible with Bandwidth Hero.

This project was structured and tested with U-Net and Yolo(v8, v11).

Input	YoloV8n (v0.1)	YoloV8s (v0.2)	YoloV11s (v0.2)

Diff

Input	YoloV8n(v0.1) / YoloV8s(v0.2)	Yolov8n(v0.1) / Yolov11s(v0.2)	Yolov8s(v0.2) / Yolov11s(v0.2)

Tune

Model	Tuning Time	Image Size	Epochs/Inter	Iterations	Fitness	Scatter Plots
YoloV8n (v0.1)	45.1h	1280x1280	100	100

Dataset versions

Property	v0.1	v0.2
Images	283	⬆️ 480
Train	249	⬆️ 420
Valid	22	⬆️ 40
Test	12	⬆️ 20
Annotations	3293	⬆️ 4832
Annotation comic	1258	⬆️ 1938
Annotation speech-balloon	2035	⬆️ 2894
Auto-Orient	Applied	Applied
Resize	Stretch to 963x1400	⬜ Fit (white edges) in 963x1400
Auto-Adjust Contrast	Using Contrast Stretching	Using Contrast Stretching
Flip	Horizontal, Vertical	Horizontal, Vertical
Rotation	❌	Between -15° and +15°
Grayscale	Applied	Applied
Exposure	❌	Between -10% and +10%
Blur	0.5px	⬆️ 1px
Noise	0.5%	0.5%

Comparison (Unet vs YoloV8 vs YoloV11)

Property	Unet	YoloV8 (v0.1)	Yolov8s (v0.2)	YoloV11 (v0.2)
Precision	0.7444	0.98424	0.96524	0.968
Recall	❌	0.95234	0.97068	0.965
Val Seg Loss	❌	0.7037	0.65568	0.69089
Val Clas Loss	0.23221	0.26816	0.31534	0.27078
Pretrained Model	❌	Yolo Nano	Yolo Small	Yolo Small
EarlyStopping	20	100	25	25
Stop Epoch	26	411	169	251
Image Set	3.882	283	480	480
Image Channels	4	3	3	3
Training Size	512x768	1280x1280	1400x1400	1400x1400
Dropout	0.2	❌	❌	❌
Kernel Size	3	3	3	3
Filter	[32,64,128,256,512]	[64,128,256,512,768]	[64,128,256,512,768]	[64,128,256,512,1024]
Artifacts	high	low	low	low

Details about Yolov8 Details about Yolov11

📝 | Cite This Project

If you use this dataset in a research paper, please cite it using the following BibTeX:

@misc{
  manga-segment_dataset,
  title = { manga-segment Dataset },
  type = { Open Source Dataset },
  author = { Ashu },
  howpublished = { \url{ https://universe.roboflow.com/ashu-biqfs/manga-segment } },
  url = { https://universe.roboflow.com/ashu-biqfs/manga-segment },
  journal = { Roboflow Universe },
  publisher = { Roboflow },
  year = { 2025 },
  month = { jan },
  note = { visited on 2025-01-24 },
}

⚙️ | Requirements

Program	Version
Python	v3.10.12

💹 | Production (proxy only)

# Install requirements
cd src
python3.10 -m venv ./python
source python/bin/activate

pip install -r requirements.txt

source python/bin/activate

# Start
python app.py

🐛 | Develop (training)

Install requirements

# Windows WSL2: https://www.tensorflow.org/install/pip?hl=en#windows-wsl2_1
# Install CUDA: https://developer.nvidia.com/cuda-downloads

sudo apt install nvidia-cuda-toolkit
sudo apt install -y python3.10-venv libjpeg-dev zlib1g-dev

Training

cd training
python3.10 -m venv ./python
source python/bin/activate

pip install -r requirements.txt
pip install --upgrade pip setuptools wheel
pip install pillow --no-binary :all:

source python/bin/activate

Yolo

# Train normally
python training/start.py --yolo --size 1400

# Look for the best result
python training/start.py --yolo --size 1400 --best

# Train on another model
python training/start.py --yolo --size 1400 --model 10

# Convert model in TensorFlow
python training/start.py --yolo --size 1400 --model 10 --convert # or only --convert without --model for latest model

# Test Model
python training/start.py --yolo --size 1400 --model 10 --test # or only --test without --model for latest model

Unet (legacy)

# Look for the best result
python training/start.py --unet --best

# Run a ready-made script
python training/start.py --unet

# Convert model in TensorFlow
python training/start.py --unet --model 3 --convert

Saving current Libraries

pip freeze > requirements.txt 

⚠️ Error Solutions

Error code: ImportError: cannot import name 'shape_poly' from 'jax.experimental.jax2tf'

Cause: This error comes from the code itself.

Solution:

jax-ml/jax#18978 (comment)

# Path: lib/python3.10/site-packages/tensorflowjs/converters/jax_conversion.py

# Remove:
from jax.experimental.jax2tf import shape_poly
PolyShape = shape_poly.PolyShape

# Add:
from jax.experimental.jax2tf import PolyShape

Error code: Wsl/Service/CreateInstance/MountVhd/HCS/ERROR_FILE_NOT_FOUND

Cause: You possibly uninstalled and reinstalled WSL/distribution.

Solution:

# List the distributions installed by running the following in PowerShell.
wsl -l

# Unregister the distribution. Replace "Ubuntu" below with your distribution name found in Step #1:
wsl --unregister Ubuntu-22.04

# Launch the Ubuntu (or other distribution) installed via the Microsoft Store

Yolo arg --best

Error:

QObject::moveToThread: Current thread (0x5a75e26f1250) is not the object's thread (0x5a75e21c6fa0).
Cannot move to target thread (0x5a75e26f1250)

qt.qpa.plugin: Could not load the Qt platform plugin "xcb" in "/home/ashu/Documents/GitHub/Manga-Segment/lib/python3.10/site-packages/cv2/qt/plugins" even though it was found.
This application failed to start because no Qt platform plugin could be initialized. Reinstalling the application may fix this problem.

Available platform plugins are: xcb, eglfs, linuxfb, minimal, minimalegl, offscreen, vnc, wayland-egl, wayland, wayland-xcomposite-egl, wayland-xcomposite-glx, webgl.

Solution:

NVlabs/instant-ngp#300 (comment)

pip uninstall opencv-python
pip install opencv-python-headless

YoloV8:

@software{yolov8_ultralytics,
  author = {Glenn Jocher and Ayush Chaurasia and Jing Qiu},
  title = {Ultralytics YOLOv8},
  version = {8.0.0},
  year = {2023},
  url = {https://github.com/ultralytics/ultralytics},
  orcid = {0000-0001-5950-6979, 0000-0002-7603-6750, 0000-0003-3783-7069},
  license = {AGPL-3.0}
}

YoloV11:

@software{yolo11_ultralytics,
  author = {Glenn Jocher and Jing Qiu},
  title = {Ultralytics YOLO11},
  version = {11.0.0},
  year = {2024},
  url = {https://github.com/ultralytics/ultralytics},
  orcid = {0000-0001-5950-6979, 0000-0002-7603-6750, 0000-0003-3783-7069},
  license = {AGPL-3.0}
}

U-Net article:

Ronneberger, Olaf, Philipp Fischer, and Thomas Brox.
"U-net: Convolutional networks for biomedical image segmentation."
In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 234-241. Springer, Cham, 2015.