zeynelsamak / captcha-decoder

A captcha solver that using machine learning and deep learning

Capthca-Decoder

A captcha solver that using machine learning(Nearest Neigbour) and deep learning(CNN,keras)

Decoding Steps

• Data Collection
  • Collecting the captcha images with the ground truth text
• Preprocessing
  • Extract the letter from the captcha images by using opencv methods (threshold > erode/dilate > countour)
  • Save the each letter to related the file - For example: All 0s to the the 0, all 9 letters to the 9 file
• Training
  • Build the CNN network by using keras or use scikit-learn to model NNeigbour classifier
  • Train the model with the extracted letters
  • When the model performance reach the optimum level, save the model weights for inference
• Testing
  • Use local file, url or base64 based captcha image
  • Extract the letter as in preprocessing stage
  • Run the inference model on each letter that extracted
  • Combine the letter prediction in order, and return the final result

Result

TODO

• Extend the explanation of the work
• Train the network without to parsing the letter. Feed whole image into the netwokr and predict the captcha text

Acknowledgement

This work was inspired by work of Adam Geitgey's medium post