Projects

Machine learning ·

Manga Bubble Cleaner.

Yet Another (Fast) Bubble Cleaner using segmented YOLOv11 for high acccuracy while maintaining inference speed.

Manga Bubble Cleaner Demo

Overview.

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!).

The main inspiration was from Not Another Bubble Cleaner. But the inference was too slow and heavy, as it uses Mask-RCNN, a quite outdated technology. We want an accurate detection, but we also want it fast. Thus we uses YOLOv11-segmented. It is quick, light, and (mostly) accurate.

This project started as a final assignment for my Machine Learning course. While brainstorming, i was thinking, “Why dont we just make a tool to help clean manga bubbles?” cause i myself is a Scanlator and i know how tedious bubble cleaning can be. So, here we are.

This is a team project consists of me, Razky, and Praja.

How It Works (oversimplified)

  1. It first detects a bubble using YOLOv11-segmented model.
  2. Then, it creates a mask from the detected bubble and applies inpainting using OpenCV to remove the bubble.
  3. Finally, it saves the cleaned image to the output folder.

What’s Next?

  • Improve detection accuracy
  • Custom mask offset
  • A GUI interface…?

This project is currently inactive and unmaintained. Feel free to fork and improve it, or whatever :)

Stack.

Machine Learning · YOLOv11 · Image Processing · Python

Next project