BallonsTranslator
BallonsTranslator copied to clipboard
dmMaze
Add MangaInpainter.
Please add the MangaInpainting method to inpaint. Since I mostly translate manga, the current methods break screentone very much, which can be avoided using the msxie92
I was thinking of adding it at the same time as EasyOCR (better result for English text). I'll probably do it next vacation if it's not already added by someone else
I would add it myself, but I don't have knowledge in pyqt and, in principle, how to embed it all into an application, I'm just learning.
There is nothing to do with pyqt, everything is in the dl folder and you already have examples
There is nothing to do with pyqt, everything is in the dl folder and you already have examples
When can you be free? I didn't understand anything how to add)
Thanks for suggesting, MangaInpainting is a fantastic work, I would like to read its paper (also its related work) before adding it to this project, perhaps later this month.
I looked at it but I saw that in fact the result was not very good either (even if still a little better than the one used). One of the problems is that comic text detector can't separate two bubbles that are quite close
I looked at it but I saw that in fact the result was not very good either (even if still a little better than the one used). One of the problems is that comic text detector can't separate two bubbles that are quite close
Can give the user this feature? Like before launches, you put a marker where there should be two, three or more clouds. Or masturbate a neuron on a hundred manga. Tell me how, I have a lot of it, I'll teach it
I was talking about EasyOCR, and I'm not good at deep learning / pyqt
I understood. And I also found out that EasyOCR does not cope well with Japanese/Chinese characters. More precisely, she often does not consider it as a text
any plan for EasyOCR ?
I'm not very convinced by its result. Its detected bboxes are actually a bit messed up compared to CTD. Also, it performs worse on manga.
Colab demo: https://colab.research.google.com/drive/1etdG7_mpoSi5ekhbek53Ex6UJO6B3hgk?usp=sharing#scrollTo=mOUcpfxMAfKV
EasyOCR
CTD
The detector is kind of more troublesome to be replaced than the recognizer. The main difference between CTD and other text detectors is that 1. ctd performs text mask segmentation and 2. textlines grouping (detect a block of text).
In order to integrate other detectors we can design some rule-based algorithms to do 1 & 2 or even utilize CTD to do them and combine the result. I don't think it is worth the effort and would rather retrain the CTD on more diverse data instead of trying other OCR or online APIs. It would be helpful if you can provide some failure cases (for analyzing not training)
comic text detector can't separate two bubbles that are quite close
I believe the main reason why it performs 2 worse on manga is that part of its training data comes from Manga109-s, which treated connected balloons as a block. Some postprocessing had been deployed to address it but things could go wrong.
By the way, are we going to start implementing this inpaint or is there no point in it?