This task is partially blocked by T144290. It can probably be started without it though.

This looks like it could be a useful basis for the OCR API tool: https://github.com/thangman22/google-cloud-vision-php

It doesn't permit a different endpoint URL, so I've raised a pull-request to add that possibility.

I've created a fork of that library, and incorporated all pending PRs (except one with this in it).

I've got a draft of a basic tool that uses the fork: it is taking a Wikisource image URL, running the image through the API, and returning the text. It doesn't restrict requests to originating from Wikisource yet, nor the size of the file. The code is in D1966.

Nope, scratch that; someone will have to tell me how to use Diffusion properly! For now (it being beer o'clock on a Friday) the code lives at https://github.com/wikisource/ws-google-ocr :-)

(I've sorted out the repository confusion. The GitHub one has been replaced by the in-house one, and the former deleted.)

Still left to do for this ticket:

• Only accept requests from wikisource.org (however it does limit image URLs to having the prefix https://upload.wikimedia.org/; is this sufficient?).
• Limit image sizes to 10 MB (I'm adding this feature to the google-cloud-vision-php library).

Actually, it look like the Vision API is limited to 4 MB per image. Can we just rely on Google complaining when a sent image is too big, and pass that error back to the user (as is already done for all errors)?

Only accept requests from wikisource.org (however it does limit image URLs to having the prefix https://upload.wikimedia.org/; is this sufficient?).

Yes, I think that should be sufficient. We just need to make sure we are not providing a way for random people to use the Google API for free (at our expense).

Actually, it look like the Vision API is limited to 4 MB per image. Can we just rely on Google complaining when a sent image is too big, and pass that error back to the user (as is already done for all errors)?

I think it would be better to limit it on the input side so that we aren't encoding 100 MB files for no reason. Let's limit it to 4 MB in that case. See https://www.mediawiki.org/wiki/API:Imageinfo for how to get this information from the MediaWiki API.

If we have files (images) > 4MB, isn't it possible for us to rescale the image so it's less than that?
Maybe using something like this.

@Niharika: yup, but actually I can't find a single proofreadpage image that is over 4 MB. Maybe for super-small-print (has anyone ever tried scanning microfinch and feeding that into Wikisource?) but it seems rare.

The image that gets send to the OCR service is already the resized version, for example: https://upload.wikimedia.org/wikipedia/commons/thumb/6/64/All_the_Year_Round_-_Series_3_-_Volume_8.pdf/page11-912px-All_the_Year_Round_-_Series_3_-_Volume_8.pdf.jpg

Individual proofreading projects can set the size of these resized images, but even when they need a large one (for example, with newspapers like this: https://en.wikisource.org/wiki/Page:The_New_York_Times,_1900-12-01.djvu/1 ) it's only 1.9 MB.

@kaldari: It's not the size of the original image that matters so much as the size of the one displayed on the proofreading page. So do you think it's too inefficient to just read that one into memory and see how big it is? I figure it's not the best, but then a call to the API would be slowish too.

I'm fixing up the Vision API interface to default to not permitting big images; but perhaps there's a better way. One possibility is also for the gadget to not even send a request for a big image (not that that would stop people from using it in other scripts).

Right, limiting to 4 MB implemented.

So now the Vision API library is pretty much done (enough), and the tool that calls it is up and running (and pointing to https://vision.google-api-proxy.wmflabs.org/ with the key that Ryan gave me).

@Samwilson: Looks like the proxy is up and running now at http://googlevision-api-proxy.wmflabs.org/. Hope that works.

Also, can you add me as a maintainer for the ws-google-ocr project?

Looks like the proxy is up and running now at http://googlevision-api-proxy.wmflabs.org/. Hope that works.

Yup, looks to be good! e.g. http://tools.wmflabs.org/ws-google-ocr/api.php?image=https://upload.wikimedia.org/wikipedia/commons/thumb/0/07/Gissing_-_The_Nether_World%2C_vol._III%2C_1889.djvu/page46-1024px-Gissing_-_The_Nether_World%2C_vol._III%2C_1889.djvu.jpg

Also, can you add me as a maintainer for the ws-google-ocr project?

Done: http://tools.wmflabs.org/?list#toollist-ws-google-ocr

@Samwilson: Looks like there's some escaping going on:
http://tools.wmflabs.org/ws-google-ocr/api.php?lang=bn&image=https://upload.wikimedia.org/wikipedia/commons/2/21/Akash_Nila_JibonAnandoDas.png

Can we have it return the raw text in Unicode (with only the normal set of escaped chars: quotation mark, backslash, control characters)?

Can we have it return the raw text in Unicode

Done.

The main code is at https://phabricator.wikimedia.org/diffusion/1966/browse/master/api.php

Reviewed.

Thanks! I've replied on that commit.

You've probably already checked it out, but there's also the general Vision API library code that this tool is relying on:
https://github.com/wikisource/google-cloud-vision-php/