Registration bot for the late torrentbytes written in February 2007. Solved their captcha to give a significant edge on the rest of the internet when hammering their registration page.
Was written by a few friends and myself as a quick hack.
The captcha had a few severe issues:
- letters aren't warped at all (constant size, etc.)
- colors are constant, making it easy to separate foreground from background
- letters are separated by at least one blank column
Captcha-solving works as follows:
-
convert the image to binary, where only blue-enough colors are black, and the rest is white (this step removes the shadows)
-
split the bitmap into letter bitmaps by looking for blank columns separating them
-
for each letter, cross-correlate its bitmap as a 1d vector with a small amount of pre-solved bitmaps. Correlation is done by using an FFT->multiplication->iFFT to perform a convolution, and the convolution with the highest maximum value is considered the match.
File details:
- c*.bmp - sample captchas, and after some processing.
- letters.dat - file with solved letters
- match_letters.py - code to cross-correlate a bitmap with the database entries to find the best match.
- ocr.py - old version of the actual bot, includes letter separation
- tb_image.py - get the text from the captcha. Does image -> b/w bitmap conversion, letter separation, calls match_letters for solving each letter
- tbbot.py - newer version of the bot in response to changes in their HTML.s