I just noticed this:



Now there's some innovation here: I haven't seen 3D letters + perspective view anywhere else, and even though it's pretty much human readable, I think it'll give bots some very hard work. I for one wouldn't know how to deal with this...

Any ideas? What do you think?

Is this new? Since last month they have had a terrible CAPTCHA system with cats and dogs appended to each letter.


Awesome AnDrEw - That's The Sound Of Your Brain Crackin'
http://www.awesomeandrew.net/

i hated that damn cat and dog system, took me like 20+ times

Yes, this is quite new... at least I noticed it only today, and I use the site quite a bit...

I agree that the cat and dog system was terrible: when there was no additional distortion the CAPTCHA was really weak (identifying the cats and dogs was easy, and the letters were of a single font, size and orientation), and after they added the warping of the letters there was no need for them, because the distortion was already very heavy...

Anyway, my only critic to this CAPTCHA is that the grid doesn't fill all the image. This way, one can detect the edges of the rectangle and inverse the perspective transformation to obtain the image straightened up... However this is not a serious flaw, because the quality is so poor that doing this process would introduce too much noise...

http://rgaucher.info/pub/rapidshare_cats.png
of what it use to be

Oh, I've recorded their whole CAPTCHA evolution all right :P

First they had this really easy CAPTCHA: fixed font, fixed letter size, fixed number of letters, very little variation of the horizontal position of the letters, fixed letter spacing, fixed letter orientation, etc. Any OCR program could beat it, but I think this is from before they were as popular as they are now, so maybe it wasn't a big issue...



Then they tried changing the font, but it obviously didn't help: seriff is good when there's some distortion to the image, so that you need to count endpoints and whatnot. It isn't the magical solution to all your problems tho...





Then they started adding noise (somebody tipped them off?), but the result wasn't much better: still too much is fixed, so very little preprocessing is necessary before an OCR program can beat the hell out of this CAPTCHA:



Then came the cats and dogs, which at first were only background noise. This probably didn't help much, because it's only two different shapes in the background, so they're easily recognizable. It does need some custom scripting tho...



Finally, the distorted letters + the cats and dogs telling you what letters to type... REALLY annoying CAPTCHA, but a lot better than the previous ones. A few weaknesses survive tho: the letters are all aligned (although they may be rotated less than 30 degrees), and the font doesn't help... those thick black borders are clues that one can use to solve the CAPTCHA rather easily (compared to google's, which uses the same principle but another more complicated font)...



Anyway, the 3D CAPTCHA in my initial post is AGES better than all of these... it may be solvable however, as the weakness I pointed out with the grid leads you to this somewhat easier problem:



EDIT: I forgot to mention that I omitted some of the less fortunate CAPTCHAs. They tried lots of things in between these, some of which were decent ideas, but most of them were so terrible they only lasted a few days...



Edited 1 time(s). Last edit at 06/29/2008 07:55PM by istari.

I think thats from http://www.ocr-research.org.ua/teabag.html. Almost the same

Seen others use it like http://www.barafranca.com

But still have no idea how to deal with it :)

Well, the last image in my previous post was my first step, and it already deals with one of the CAPTCHA's more important features (i. e. the perspective).

The next step would be to analyze the amount of black and white in non-data areas (which is quite constant for small rectangles), and then discard all the parts of the image which have that same amount. That leaves the areas with too much black or too much white, which are exactly the borders of the letters.

I still have to implement this, tho, so I don't know how well it will work...