YomiNinja
YomiNinja copied to clipboard
matt-m-o
OCR mistakes
First, let me say thank you for the amazing and immensely helpful tool! I just used it for a ~20h playthrough and I'm convinced it has saved me several hours of manual dictionary lookups. With that said, the OCR is not perfect and it makes a lot of relatively simple mistakes. Not sure how much control you have over the OCR engine, but in case there are some knobs to turn, take a look at these. I tried changing the OCR engine but it didn't seem to help much.
Thank you! I'm really glad to hear that the tool is proving helpful for you!
The current OCR engine (PaddleOCR) has limitations with some kanji not included in the training data, like [ 瞥 恫 怯 恍 聳 ]. To improve recognition for characters like [ 卑 ] and others, you can try increasing the "Maximum image width" slightly (values higher than your screen width might not take effect) and enabling "Invert image colors" for white text.
In version 0.6, I integrated the Google Cloud Vision OCR engine, which doesn't exhibit the mentioned issues. PaddleOCR now has more parameters for tweaking and that might help a bit too.
I'm also in the process of integrating Manga OCR, which seems promising. It's capable or recognizing a larger variety of fonts and kanji and didn't fail on your examples. I'm working on the issue of wrapped lines for the next update as well.
There is also the issue I open to PaddleOCR to improve jouyou kanji recognition, a +1 on that could be useful but I’m out of the country and have terrible data coverage right now, can you please share the link?El 31 ene 2024, a las 20:00, Matheus M. @.***> escribió: Thank you! I'm really glad to hear that the tool is proving helpful for you! The current OCR engine (PaddleOCR) has limitations with some kanji not included in the training data, like [ 瞥 恫 怯 恍 聳 ]. To improve recognition for characters like [ 卑 ] and others, you can try increasing the "Maximum image width" slightly (values higher than your screen width might not take effect) and enabling "Invert image colors" for white text. In version 0.6, I integrated the Google Cloud Vision OCR engine, which doesn't exhibit the mentioned issues. PaddleOCR now has more parameters for tweaking and that might help a bit too. I'm also in the process of integrating Manga OCR, which seems promising. It's capable or recognizing a larger variety of fonts and kanji and didn't fail on your examples. I'm also working on the issue of wrapped lines for the next update.
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you are subscribed to this thread.Message ID: @.***>
