Luogu Captcha Predict

Introduction

Recognize Luogu Captcha using an AI model.

Usage

1. Ensure you have TamperMonkey or another UserScript manager installed in your browser.
2. Install the UserScript by downloading the file predict.user.js.

Model and Dataset

The Model and Dataset are hosted on HuggingFace.

Current Model Information

• Training data is located in the data folder. Note: The folder contains 1000 files, so exercise caution when opening it in your browser!
• Training history visualization:

Data Generator

The data generator consists of generate.php and its Python wrapper generate.py.

• generate.php
  • Without arguments: Generates a captcha, outputs the captcha answer to stdout, and saves the image as captcha.jpg.
  • With two arguments (tot, seed): Both arguments must be integers. The program generates tot images using the random seed seed, concatenates all image data, and outputs it to stdout. Each image is formatted as follows:
    • First 2 bytes (len): Length of the image data
    • Next 4 bytes: Captcha answer
    • Next len bytes: Binary image data
• generate.py
  • Requires three arguments: TotalImages and WorkersCount. It generates TotalImages image batch files in the data directory, formatted for HuggingFace (data/luogu_captcha_dataset) and TensorFlow (data/luogu_captcha_tfrecord).

Model Training

The script train.py trains the model using TensorFlow with data from the data/luogu_captcha_tfrecord folder. The trained model is saved as models/luoguCaptcha.keras.

Predicting Captchas

The script predict.py is used for captcha prediction.

• With one argument (port): Starts an HTTP server on the specified port, providing a single API endpoint.
  • URL: /
  • Request Method: POST
  • Request Body: JSON in the following format:

    {
        "image": "base64 encoded image file"
    }

  • Response Body: JSON in the following format:

    {
        "prediction": "the captcha answer"
    }

License

This project is licensed under the terms of the GNU General Public License v3.0.