🎈 Manga/Comics Speech Balloon Detection using YOLOv12

📌 Overview

This project is designed to train a YOLOv12 model to detect speech balloons in manga and comic images. It includes:
✅ Pre-processing scripts for dataset organization and label formatting.
✅ A training script for model development.
✅ An inference script for applying the trained model to new images.

📂 Directory Structure

backup\
└── balloon_training\
    └── weights\
        ├── best.pt
        └── last.pt
cfg\
└── balloon_yolo_config.yaml
dataset\
├── images\
│   ├── train\
│   ├── valid\
│   └── test\
└── labels\
    ├── train\
    ├── valid\
    └── test\
scripts\
├── inference.py
├── round_labels.py
└── split_dataset.py
README.md
train.py
requirements.txt

📄 File Descriptions

🛠 cfg/balloon_yolo_config.yaml

This YAML configuration file defines the dataset paths and class names for training. It includes:
🔹 Relative dataset paths.
🔹 Directories for training, validation, and testing images.
🔹 The label class name.

path: "../dataset" # If you face issues, replace this with the absolute path to your dataset folder.
train: images/train
val: images/valid
test: images/test

names:
  0: balloon

---

🔍 scripts/inference.py

Purpose:
Runs inference on a folder of images using a trained YOLO model.

Usage:

python scripts/inference.py --weight <path_to_weights> --img_folder <path_to_images> --output_folder <output_folder_name>

How It Works:
✅ Loads the YOLO model from the specified weights.
✅ Iterates through images in the given folder.
✅ Resizes images based on height and runs inference.
✅ Saves output images with bounding boxes in the output folder.

---

🏷 scripts/round_labels.py

Purpose:
Ensures label consistency by rounding numerical values to four decimal places.

Usage:

python scripts/round_labels.py <label_directory>

How It Works:
✅ Scans the specified folder for .txt label files.
✅ Rounds each numerical value (except class index) to four decimal places.
✅ Overwrites the original files with rounded values.

---

📂 scripts/split_dataset.py

Purpose:
Splits the dataset into training, validation, and test sets based on defined percentages.

Usage:

python scripts/split_dataset.py <dataset_path> --train_pct 70 --valid_pct 20 --test_pct 10

How It Works:
✅ Checks that percentages add up to 100.
✅ Creates the required subdirectories for images and labels.
✅ Randomly shuffles images and assigns them to train, validation, and test sets.
✅ Moves corresponding label files along with the images.

---

🎯 train.py

Purpose:
The main script for training the YOLO model on the speech balloon dataset.

Usage:

python train.py --model <model_file> --data cfg/balloon_yolo_config.yaml --epochs 1000 --batch 16 --imgsz 640 --project backup --name balloon_training --cache ram

How It Works:
✅ Detects GPU availability for optimized training.
✅ Loads the YOLO model with pre-trained weights.
✅ Trains using specified epochs, batch size, image size, and caching method.
✅ Saves the best and latest weights in the backup directory.

---

🛠 Installation & Requirements

This project uses requirements.txt for dependency management. Install all required packages with:

pip install -r requirements.txt

Key dependencies:

• torch
• ultralytics
• opencv-python
• pillow

---

🔄 Usage Workflow

1️⃣ Dataset Preparation

Use scripts/split_dataset.py to split your dataset into training, validation, and test sets.

2️⃣ Label Processing

Run scripts/round_labels.py to ensure that all label values are rounded for consistency.

3️⃣ Model Training

Execute train.py with the appropriate arguments to train your YOLO model.

4️⃣ Run Inference

After training, apply scripts/inference.py to detect speech balloons in new images.

---

ℹ️ Need Help?

If you need more details about project setup, script functionalities, or troubleshooting, feel free to ask! 🚀