|
--- |
|
languages: |
|
- English |
|
tags: |
|
- recaptcha |
|
- captcha |
|
- hcaptcha |
|
- prediction |
|
- torch |
|
- efficientnet |
|
- keras |
|
- coreml |
|
- tensorflow |
|
- webapp |
|
- python |
|
--- |
|
|
|
* Captchot Images |
|
|
|
** Captcha Images Prediction ** |
|
|
|
This model is the images version of the Captchot project, a from-scratch experimental project containing two models to predict image based captchas and text-to-image based captchas |
|
|
|
** Dataset ** |
|
|
|
The dataset used is a dataset of 300k+ images taken from real-world captcha dumps (mainly reCaptcha 1 and 2 and hCaptcha plus others found to be used online) divided in categories (such as trains, cars, stairs...). |
|
|
|
The complete list of labels can be found in Structure.png. |
|
|
|
** Files ** |
|
|
|
You can find the training overall informations in training_info.png, the .zip structure in Structure.png and the demo webapp screenshot in Webapp.png. |
|
|
|
Captchot_Images.zip contains the exported models for different frameworks: tensorflow, onnx, keras, a webapp demo, a python ready to use script and coreml . |
|
|
|
** Model ** |
|
|
|
The model has been trained with EfficientNet at 1000 iterations and early stop to avoid overfitting. The full network has been trained. |
|
|
|
** Useless info ** |
|
|
|
The training has been done on a M1 Pro Apple Silicon chip and took approximately 20h to fully train (plus the dataset importing phase). |
