Let's talk captcha breaking (GOCR,TESSERACT,etc)

[jb2008]

Anyone in this thread who can code Xrumer captcha DLL's, please contact me.

 

[goodbuyer]

I use python pil, to get some easy captchas solved, not hotmail, not recaptcha. First I tried using tessaract out of the box and never tried to train it, I thought it would be too much work, instead I used other methods like neural networks, but some easy captchas still got me troubled because of rotation.

The algorithms that I used for the cleaning part are.

Binarization algorithms like otsu method
mean binarization
image algebra functions (erosion, dilation...)
some noise cleaning functions and dot cleaning functions (simple stuff)

Hope this helped

 

[SalonSpaSource]

How about a team of people working on Captcha 1000 Captcha for 1$
If interested PM me.

 

[blakamia]

I've also been messing with captcha cracking. I've broken Elgg and ExpressionEngine captcha with 80%+ accuracy using my own engine (for ExpressionEngine I use Tesseract after cleanup to recognize words). I'm currently working on Recaptcha using only my own engine (no Tesseract), and am finding it easier than expected. If you clean up images correctly you will find that recaptcha reuses the letters font/size in the code word a lot, so you can easily recognize it

Aside from that I haven't tried any other captchas that are worth mentioning. Just simple stuff in directory submissions for example.

 

[kelvin.thechamp]

i use peguass smart scan engine sdk and working on recaptcha ocr currently having 30% success
making it more accurate day by day i code things in c# and delphi
if any help require very happy to give my input !

 

[mikeybobikey]

Great thread!!

I have extensive knowledge in this area!! Im the author of an <unnamed not to be mentioned here on bhw > bot!

Ive been working on this for roughly a year now. I use a combination of ImageMagick for quick pre-processing. Additionally, if the pre-processing does not yield good results - and this is the case with most moderately complex captchas, then I also work on the image in VB, where I have written many captcha specific and general pre-processing algorithms.

Once that pre-processing is complete, I then feed it to either tesseract or gocr. Both of these have their strengths and weaknesses and both can be trained. tesseract seems to be the best choice in 90% of the cases. Training tesseract can be a little tricky especially if you follow the docs... it wont get u very far

id be interested to maybe share algorithms and such. The latest accomplishment is a very fast character derotation algorithm that Im using now. It derotates almost perfect everytime.

Ive also got a killer list of bookmarks on captcha solving... beleive it or not .. most of these algo's are already out there just waiting to be ported to your language of choice.

Best,
Mikey

 

[The Data Guy]

Using tesseract I coded an ocr server that processes images and returns the text from the image. Its currently in its beta stages and I will post back here with the sucess rate.

 

[johndea]

Using tesseract I coded an ocr server that processes images and returns the text from the image. Its currently in its beta stages and I will post back here with the sucess rate.

How is it going? Please report back.

 

[The Data Guy]

I am going to post the source code in a few

 

[The Data Guy]

I've uploaded the source simplyy because I feel it might help others. I havent gotten around to testing it, because i am probably going to develop a forum poster and use this ocr setup in it. http://www.blackhatworld.com/blackhat-seo/visual-basic-net/330191-c-client-server.html#post3015305

 

[Rsquared]

After readin this thread, I tried a little experimenting myself and well...got carried away.

Image here(I'm not allowed to post links/imgs yet):
img84.imageshack.us/img84/1153/screenshot20110720at455.png


I managed to separate/extract POF's captcha into single images, but tesseract was NOT doing a good enough job for me, so I created an alphabet folder which I used to compare against each extracted image. There's a few letters that tend to mess up, but it's working pretty well.

 

[cody41]

Nice, so are you initially using tesseract to box out the letter elements and then using something else to ocr?

 

[daymond]

I've also been messing with captcha cracking. I've broken Elgg and ExpressionEngine captcha with 80%+ accuracy using my own engine (for ExpressionEngine I use Tesseract after cleanup to recognize words). I'm currently working on Recaptcha using only my own engine (no Tesseract), and am finding it easier than expected. If you clean up images correctly you will find that recaptcha reuses the letters font/size in the code word a lot, so you can easily recognize it

Aside from that I haven't tried any other captchas that are worth mentioning. Just simple stuff in directory submissions for example.

have you made some working recaptcha solver that can work in multi thread mode? what is the speed of cracking and how much resources does it consume?

are you guys also using maybe neural networks to crack captchas?

 

[Rsquared]

Nice, so are you initially using tesseract to box out the letter elements and then using something else to ocr?

I actually created my own method of boxing out the letters

 

[goodbuyer]

After readin this thread, I tried a little experimenting myself and well...got carried away.

Image here(I'm not allowed to post links/imgs yet):
img84.imageshack.us/img84/1153/screenshot20110720at455.png


I managed to separate/extract POF's captcha into single images, but tesseract was NOT doing a good enough job for me, so I created an alphabet folder which I used to compare against each extracted image. There's a few letters that tend to mess up, but it's working pretty well.

That should be very very easy to tessaract after cleaning, characters are very separate from each other, edges are clean no hair no noise.

I should do one of two things.

1) Analize the image in a vertical way so to separete de line of circles and symbols from the letter under.

2) Get a circle triangle square recognizer so to clean the iamge from these figures. This can be done in Matlab out of the box.

 

[roknimam]

Btw I wonder did you guys sow one guy use google google to brake any type of capcha some Russian guy already made software for that. I think I came across that video on jdownloader forum, any thoughts how we can abuse that, he also mention there is no limitation at the moment but google will catch up with time.

 

[raginrix]

Btw I wonder did you guys sow one guy use google google to brake any type of capcha some Russian guy already made software for that. I think I came across that video on jdownloader forum, any thoughts how we can abuse that, he also mention there is no limitation at the moment but google will catch up with time.

Do you have a link to that? And he's using Google to break Google's captcha? Now that's pr0

 

[The Data Guy]

In order to remove the squares and triangles from the image use Aforges quadrilateral
filter.

 

[The Data Guy]

SO any updates

 

[ukneeq]

So does anyone have the training data or trained eng files for tesseract that can be used for reCaptcha?