cagdass / stars-captcha-stuff

Captcha solver for Bilkent's Stars

Stars Captcha Stuff

Before getting started

This project is incomplete, I have to return a session with cookies after sending the correct captcha. However, if you're in a hurry, just fire your browser go to any page in Stars that requires captcha before letting you in, enter the captcha and check your cookies.

Use the key PHPSESSID and its value found to send as cookies in your application and it'll work like a charm.

Getting started

Captcha solver for Bilkent's Stars

Help needed: Although the captcha is correctly read, there is probably an issue with sending it as a string. The reasons may be related to the cookies, user-agent or some other dipshit. This part remains unsolved.

Usage

# First, clone the repository.
git clone https://github.com/cagdass/stars-captcha-stuff

# There are a few Python libraries you may need to install,
# can be done while getting errors.

# If you'd like to reproduce the letters, else skip to *$@*:
mkdir images
python save_images.py # Saves 100 captcha images from Stars, this number can be modified.
python get_digits.py # Asks the user to label each character extracted from the image, labeling part can be modified.

# *$@*:
python main.py

Improvements

The letters and digits from the captchas have already been extracted, though some have a noise of a pixel or two.

However, although each letter and digit has the same height in the letters folder, the image width of the letters and digits is not standardized, each varies with regard to the characteristics of the character - the image with w is wider than v.

• To do: Have a fixed width, add extra pixels to images. If the image width w and height h, and the fixed width is W, add 0 pixels on the left and the right, the sizes of these blocks should be h x w/2 on each side, if w % 2 != 0, the right side can have an extra vertical line h x 1.
• When I gave it a quick run, the program correctly classified 82 characters out of 100 characters. However, the success rate was only 55%. Since all 5 characters in the captcha image must be correctly classified in order to get access, this needs to be improved. One idea is to extract multiple instances of characters, say 5 images each character, and run a KNN classifier.