From my searches through the depths of the internet, I've come to the conclusion that it would be quite difficult to come up with a Japanese OCR for vobsubs, because apparently it's different from DVD to DVD.
Granted, I've never actually gone through and tried to do it, but that's what people seem to say.

Plus, the OCR software I've stumbled upon only do ASCII characters, and the ones that don't...for instance, I had 浜村, and for 浜 it gave me the water-radical, the "axe," and the "animal feet" as possible letters. Perhaps OCR programs were programmed with Heisig in mind?

From what I understand about OCRing (which isn't much, mind you...next to none, actually) is that it will scan the picture for characters, "recognize it" as a character/word, and save it to a text file.

Doing this kanji-catchpa thing, would be equivalent to making a giant database of kanji/words that show up in vobsubs, so then the OCR can recognize them more effectively?

OR! it just occurred to me that the catchpa could actually BE the OCR.
But, with my experience (ie. 浜), the word/characters it shows you don't always match up with actual characters. I've seen some programs that allow you to fix this, but I haven't used any of them, so I'm afraid I don't know much about it.

ropsta wrote:

*Sees word captcha and immediately bails*

I used to hate captchas. Once they started using reCAPTCHA everywhere, I became okay with it. Also, the reCAPTCHA words are more entertaining and fuzzier than what I remember in the past, in terms of what you can mistype and still pass.

So I say, let's make a 'what's this Japanese text?' reCAPTCHA game, but with words taken from vobsubs! Maybe have an OCR program provide a best guess, then in the context/translation as presented as a game, users can do the rest. It's fun. Yes, fun.

Annotating books with audio in Transcriber is also fun and great listening/reading practice, and you end up with interactive books for use in the Kage Shibari reader! It's fun, damn it!

Last edited by nest0r (2010 April 05, 3:13 am)

nest0r wrote:

Did you read what recaptcha does? ;p

I know what recatchpa does, i've known the general idea for a while, just never looked into exactly how it works.

But that's not my point, nor does it answer my concerns. Even if someone were to develop a Japanese catchpa for vobsubs, it still doesn't solve the problem that I've had with OCRs in the past.
What happens if you get instead of 泣、and the next character it gives you is 立.
you could end up with "彼女は止まらなく三水立いた"

Jarvik7 wrote:

How exactly would you end up with that? Users entering the captcha would need to type in 泣. It isn't OCR...

More specifically, each word that cannot be read correctly by OCR is placed on an image and used as a CAPTCHA. This is possible because most OCR programs alert you when a word cannot be read correctly.

Doesn't it work like
1. OCR goes through, and gets things wrong
2. Images are made of things that are incorrect
3. Users type in what is wrong in order to correct them
4. Yay, we can read now!

The problem that I see occurs in between step 1 and 2. It goes through and sees 泣いた. When I was OCRing 浜村, I was prompted with the　三水, as opposed to 浜.
So, if an image is made of the 三水 as opposed to 泣, how are the users going to know to type in 泣、as opposed to 河, 浜、or even 水 for that matter.

I must be missing something somewhere...

@Asriel - As the quote says, it doesn't make an image of the incorrect item. It takes what it's not sure about or doesn't know and presents that image to the user, alongside something it does know. The source image. Then, provided the user gets the one it does know correct, it assumes the same for the other, building a more and more certain aggregate over bazillions of attempts. From there I have no idea what happens.

So you don't have to worry about the user looking at the incorrect image, just the unknown but accurate one, should anyone be able to decipher it.

For learners of Japanese rather than experts, doing vobsubs, I suppose a different logistics would be in order.

For vobsub, I only did it once, ages ago, bt it seems via a d-addicts tutorial I glanced through, that it's mostly user-oriented anyway. The computer simply presents the images to you and you input it yourself, teaching the system? The tutorial recommended using the IME input pad for stuff you don't know. -- http://www.d-addicts.com/forum/viewtopic.php?t=16017 (bottom 'The Alternative')

As an aside, you speak from this single problem, so I wonder, how many different Japanese OCRs have you used, Asriel? Did you read that thing I didn't post in a totally innocent thread? That was from a low quality source, required perhaps 2 corrections a line if memory serves, and was easily fixed via human-looking-at-source-image.

At any rate, between the possibilities of providing users with source images and context, different input methods, compounds, translations, etc., I think there's lots of possibilities there to 'crowdsource' or whatever rather than relying on machine vision. One movie's vobsubs vs. 200 people OCRing it w/ human vision, etc.

There's also Transcriber's style... What if you have the film audio playing, the vobsubs in front of you, and that way you have the 'readings' that you can add phonetically, pausing and inputting kanji as you go along? Or something. Anyway, I'm just brainstorming. ;p

I didn't even go into how reCAPTCHA refines it to combine the two, because the OCRs I've used, I didn't know there was a place where it told you when it was spitting out something even though it wasn't sure. If I could tell it not to do that and instead to simply mark that image as unknown, that would be good...

Also, it did answer your concerns and was your point, because even if you knew what reCAPTCHA does, as you said, you didn't know exactly how it works, even though I went out of my way to post the quote in question that answered your concerns. So don't get snarky with me! ;p

Also, can't we type in Anki? Maybe people can start typing up their vobsub'd film cards/decks made in subs2srs, and donate the results?

Last edited by nest0r (2010 April 05, 2:12 pm)