Fast Bubble Cleaner

Bubble Cleaner untuk manga menggunakan model YOLOv11-segmented, untuk inferensi yang lebih cepat dan ringan, selagi mempertahankan akurasi deteksi yang tinggi. GPU tidak diperlukan (CPU saja cukup!).

Bubble Cleaner for manga using the YOLOv11-segmented model, optimized for faster and lighter inference while maintaining high detection accuracy. A GPU isn't needed (Just CPU is enough!).

Usage

Untuk pemakaian, silakan download versi yang sudah di-Compile untuk CPU-only (Windows, .exe) di halaman Release. Download file .zip dibawah asset, ekstrak, lalu jalankan exe. Tunggu hingga teks yang meminta memasukkan input path muncul.

For usage, please download the pre-compiled CPU-only version (Windows, .exe) from the Release Page. Download the .zip file under the "Assets" section, extract it, and then run the .exe. Wait for the prompt asking for the input path to appear.

Usage Example

1. Create input and output directory.

2. Run program.

3. Enter input and output directory, and the Manga Panel to be processed.

4. Wait until finished, done.

Note

This model isn't fully optimized for bubble detection and refining, so use it with a grain of salt. Unaccuracy is expected, so use it just as a helper, not to remove Cleaner role entirely.

Class Detection Features

• Ellipse / Bubble default
• Polygon
• Square / Narrative Dialogue
• Thorns / Shout (kurang dataset)

Demo

Training Model

Training dilakukan di Google Colab menggunakan gpu T4, dengan epoch 200 (berhenti di epoch 54 karena tidak ada peningkatan).

Confusion Matrix

Sebaran Label/Dataset

Hasil Akhir

iterasi epoch	train/box_loss	train/seg_loss	metrics/precision(B)	metrics/recall(B)	metrics/mAP50(B)	metrics/mAP50(M)
54	0.2383	0.28144	0.86595	0.80737	0.82864	0.82972

Untuk hasil iterasi epochs secara lengkap dapat melihat pada link ini: https://drive.google.com/file/d/1lJIBTGtpXOicD8E6ZnQcW2NUkYRSgfY4/view?usp=drive_link

To-Do List

• Increase Dataset Variations
• Improve Post-processing, especially for joined bubble

Run Locally

Clone the project

  git clone https://github.com/faralha/Bubble-Cleaner.git

Go to the project directory

  cd Bubble-Cleaner

Configure path

This file was imported from google collab, and all the training dataset and models were assumed stored in Drive. Change this accordingly.

For Training: Configure Dataset Path
For Inference: Configure Models Path

Run Notebook

• train.ipynb for Training Purposes, and
• inference.ipynb for Main Usage