captcha_break

Why we included this project

captcha_break is a Keras walkthrough that teaches a convolutional network to read four-character text CAPTCHAs, predicting each position with its own classifier. The data pipeline is its best part: instead of scraping or hand-labeling images, it synthesizes unlimited training samples on the fly using the captcha library and a Keras Sequence generator. That makes it a practical starting point for developers exploring sequence-recognition OCR, with runnable Jupyter notebooks that walk through VGG-style blocks, batch normalization, Adam with callbacks, GPU memory handling, checkpointing, and loss curves. One caveat: the code targets TensorFlow 1.13-era APIs, so expect to update imports and fit_generator calls before it runs on modern frameworks. A PyTorch variant of the notebook is included for anyone who prefers that stack.