kha-white / manga-ocr

Optical character recognition for Japanese text, with the main focus being Japanese manga
Apache License 2.0
1.74k stars 89 forks source link

Struggles with white letters with a black outline/"bubble letters" #81

Open ngoomie opened 3 months ago

ngoomie commented 3 months ago

The manga Yakusoku no Neverland seems to have a lot of white text with a black outline in it, and manga-ocr seems to struggle with this to varying degrees. One example: Output:

日常に酒む「意図」

This is actually a bit "nicer" of an example, one it's not struggling with as bad as some of the others. I'll have to backtrack and find the other worse examples once I'm done this chapter, but I just wanted to open an issue before I forget again.

Do note that I'm using mokuro for this, so the screenshots I'm submitting as examples here aren't the actual images that were processed, so maybe the output will be a tad different?

ngoomie commented 3 months ago

More:

Output:

『田』

Output:

そして周りを取り囲む『認

Output:

いいけど、

Output:

MoeMonsuta commented 2 months ago

Invert the images in your favorite graphics editor and it works flawlessly.

2024-08-30_190149

宣戦布告