This is a simple CAPTCHA breaker I wrote. It only breaks a specific type of captcha. Here is some proof that it works! :D
python predict.py to load the model and run some predictions. It requires that there are images in the tests/ directory to load from. You can use the included script to download images.
I tried many different things over the course of this to get it to work. My first attempts were more methodical in operation. I tried some custom image transformations, ultimately passing it into the tesseract library. This gave limited success, but it wasn't great.
The current model is actually 4 single-character models trained to recognize a specific character position. Each single-character model is a an 8-layer Convolutional Neural Net, with 3 Convolutional layers, each with 32 filters (except the last, which has 64).
I trained the model on 100000 captcha images. After downloading and manually solving 200 samples, I realized it would be infeasible to get the data I need. So, to compensate, I used the custom image transformations I had developed earlier to extract the background image from this captcha. From there, I used PIL to auto-generate captcha images, which were close enough to be used as training data. If you can generate your own data, you can train to arbirary accuracy.
brew install tesseract
pip install -r requirements.txt
Then, run ./create_tests.sh to download some captchas. You'll have to manually solve them. Solutions should be in the form: <solution>.jpeg. Example: fv8w.jpeg.
This is a work in progress. Released under the MIT License.