Captcha breaker

Based on vkCaptchaBreaker

Notable Changes:

• removed mobile network
• simplified to the working minimum
• introduce acc metric
• packed into tf serving

Setup

For better experience on Windows I strongly recommend using TF via docker deployment.

Personally, I've used TF with jupyter bundled for training and testing this model. That is also included into docker-compose

Training

1. Create train/ folder in repo root

2. Put your dataset in it

3. Just run training.py

   Optional

4. Run cast_to_inference_model.py after training to get inference model

5. Deploy via tf serving

My results

Trained on dataset of ~1k manually labeled captcha images.

Achieved acc >0.95, for better results bigger dataset required.

Note: trained mostly on 4-5 symbols images, acc will drop significantly when recognizing 6-7 symbol images. However, it should perform well in case you have big enough dataset with captcha of various size.

Sources:

• vkCaptchaBreaker
• OCR model for reading Captchas