langningchen / luoguCaptcha

Recognize Luogu Captcha with AI model

Luogu Captcha Predict

Introduction

Recognize Luogu Captcha with AI model

Usage

1. Make sure you have Tampermonkey or other userscript managers installed in your browser.
2. Install the userscript by file predict.user.js.

Develpment

Current model info

• Training data is in the folder data. Caution: the folder has 1000 files in it, be careful when opening it in your browser!
• Training history image:

Data generator

The font folder contains the font files for generate the images. The data generator is generate.php, and the generate.py is a wrapper of the php file.

• generate.php
  • When there is no arguments passed to the file, generate a captcha, output the captcha answer to stdout and write image to file captcha.jpg.
  • When there is one arguments (tot) passed to the file, the parameter must be an inteager. The program will generate tot images, connect all image data directly and output to stdout. Each image is output in the following format:
    • First 2 bytes (len): the length of the image data
    • Next 4 bytes: the captcha answer
    • Next len bytes: the binary image data
• generate.py
  • Require three arguments: BatchNumber WorkersCount BatchSize. It generates BatchNumber image batch files in the data directory, each file contains BatchSize images, working with WorkersCount subprocesses. The batch files are named as Data_([0-9]+)\.pkl

Model training

The file train.py trains the model in TensorFlow with the data generated in the folder data. The output model files is luoguCaptcha.keras, the output train history file is trainHistory.png.

The detailed model is shown below:

• Epochs: 8
• Layers:

Name	Type	Output shape
input	input	(35, 90, 3)
con1	Conv2D	(33, 88, 32)
pol1	MaxPooling2D	(16, 44, 32)
con2	Conv2D	(14, 42, 64)
pol2	MaxPooling2D	(7, 21, 64)
con3	Conv2D	(5, 19, 128)
pol3	MaxPooling2D	(2, 9, 128)
flat	Flatten	(2304,)
drp	Dropout	(2304,)
des1	Dense	(1024,)
des2	Dense	(CharsPerLabel * CharSize,)
output	Reshape	(CharsPerLabel, CharSize)

In this model, the CharsPerLabel is 4, CharSize is 128

Predicting captcha

The file predict.py predicts the catpcha.

• When there is no arguments passed to the file, predict the image captcha.jpg and output the answer in the stdout.
• When there is one argument (port) passed to the file, run an HTTP server on port port, and it only has one API.
  • URL: /
  • Request Method: POST
  • Request Body: A JSON as the following format

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

  • Response Body: A JSON as the following format

    {
        "prediction": "the captcha answer"
    }

License

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