tensorflow-captcha

Training Constraints

1. Only support 4 digits as both correct answers and contents in images
2. Only support an alphanumeric combination of lowercase English letters and numbers.

Enviroment Set-up

Using Anaconda to set up virtual dev environment is recommended.

Specified Version Requirements

1. Python 3.8
2. TensorFlow 2.7

Model Training Steps

1. Collect catpcha-like images and rename them with correct answers. For example, if the image says '5dc8' in your eyes, it shall be renamed 5dc5.jpeg.
2. Put a major portion of images in a root directory named training for training.
3. Put another batch of images in another root directory named validation for validation.
4. Run python train.py, a prompt will ask how many epochs do you want. Insert an integer.
5. Run python predict.py, images in the validation folder will go through the model.
   • A series of predictions will show up in your terminal.
   • At the end, a percentage will tell what is the preciseness of current model.

Personal Training Experience

After labeling over 700+ images, after a certain amount of epochs, the correct percentage could have reached around 70%. The more images, the better the result is.

Concepts

Inspired by Andy Wu's Note