A deep learning project using Keras to build convolutional and recurrent neural networks for high-accuracy captcha recognition.
captcha_break is an open-source deep learning project designed to automatically recognize and decode captcha images. It uses convolutional neural networks (CNNs) and recurrent neural networks (RNNs) with CTC loss to achieve high accuracy on alphanumeric captchas, solving the problem of automated verification code bypass for research and educational purposes.
Machine learning practitioners, researchers, and students interested in computer vision, OCR, and deep learning applications for security or automation tasks.
Developers choose this project for its well-documented, end-to-end implementation, high accuracy (over 99%), and practical focus on real-world captcha challenges, including distinguishing similar characters like '0' and 'O'.
验证码识别
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Achieves over 99% accuracy on alphanumeric captchas, with specific examples demonstrating successful distinction between challenging characters like '0' and 'O'.
Provides a complete workflow from on-the-fly data generation using a custom Keras Sequence to model evaluation, making it practical for learning deep learning applications.
Implements both CNN for character classification and RNN with CTC loss for sequence recognition, showcasing modern techniques like bidirectional GRUs and GPU acceleration with CuDNNGRU.
Integrates callbacks for early stopping, model checkpointing, and CSV logging, along with multi-processing support for efficient training workflows.
Relies on TensorFlow 1.13.1 and specific library versions from 2019, which are obsolete and may cause installation issues or lack support for newer hardware and APIs.
Only handles fixed-length alphanumeric captchas (digits and uppercase English letters), not adaptable to other formats like symbols, lowercase letters, or variable lengths without code modifications.
The README is primarily in Chinese, with some technical terms in English, which may limit accessibility for non-Chinese speakers despite the detailed code examples.