ulozto-captcha-breaker

Deep learning model using Tensorflow that breaks ulozto captcha codes.

Algorithm used will be described in a standalone document.

How to use pretrained model in your project

Prerequisities

Packages

• numpy~=1.18.3
• tflite_runtime~=2.5.0

You need to install Tensorflow Lite Runtime with the correct version depending on your operating system and instruction set. It can be found here: https://www.tensorflow.org/lite/guide/python.

Model specification

• Input shape: (batch_size, height, width, 1), where height = 70, width = 175
• Output shape: (batch_size, number_of_letters, number_of_classes), where number_of_letters = 4 and number_of_classes = 26

Note that it takes grayscale images as the input. RGB images therefore have to be converted.

Steps

1. Go to latest release and download binary files

2. Instantiate the tflite interpreter. For that you're going to need TFLite model. You can find it in the release binary files.

   • PATH_TO_TFLITE_MODEL is path to directory containing the neural network pretrained model

   import tflite_runtime.interpreter as tflite
   interpreter = tflite.Interpreter(model_path=PATH_TO_TFLITE_MODEL)

3. Normalize image to 0..1 interval. If it already is, skip this step.

   img = (img / 255).astype(np.float32)

4. Predict using following code

   # convert to grayscale
   r, g, b = img[:, :, 0], img[:, :, 1], img[:, :, 2]
   input = 0.299 * r + 0.587 * g + 0.114 * b
   
   # input has nowof  shape (70, 175)
   # we modify dimensions to match model's input
   input = np.expand_dims(input, 0)
   input = np.expand_dims(input, -1)
   # input is now of shape (batch_size, 70, 175, 1)
   # output will have shape (batch_size, 4, 26)
   
   interpreter.allocate_tensors()
   input_details = interpreter.get_input_details()
   output_details = interpreter.get_output_details()
   interpreter.set_tensor(input_details[0]['index'], input)
   interpreter.invoke()
   
   # predict and get the output
   output = interpreter.get_tensor(output_details[0]['index'])
   # now get labels
   labels_indices = np.argmax(output, axis=2)
   
   available_chars = "abcdefghijklmnopqrstuvwxyz"
   
   def decode(li):
       result = []
       for char in li:
           result.append(available_chars[char])
       return "".join(result)
   
   decoded_label = [decode(x) for x in labels_indices][0]

   • np for numpy

How to train your own model

1. Install environment Following script creates new virtual environment. You can of course use global environment instead. All following section's scripts are expected to be executed from repository's root directory.

   git clone https://github.com/JanPalasek/ulozto-captcha-breaker
   cd "ulozto-captcha-breaker"
   
   # create virtual environment
   python -m venv "venv"
   
   source venv/bin/activate # or .\venv\Scripts\activate.ps1 in windows Powershell
   python -m pip install -r "requirements.txt" -e .

2. Obtain dataset of captcha images and store it to directory out/data. Images are expected to be named according to captcha displayed in the image.

   E.g.

   

   This captcha image is expected to be named e.g. ABFD.png, abfd.png (if we don't care about case sensitivity) or e.g. ABFD_{UUID4 CODE}.png (to distinguish different images for same captcha letters).

   This project contains a way to generate captchas yourself using captcha Python package using script bin/simple_captcha_generate.py. You can run it in a following manner

   python bin/simple_captcha_generate.py --height=70 --width=175 --available_chars="abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ" --captcha_length=6 --dataset_size=10000

   Some of notable parameters are:

   • available_chars - list of characters that will be generated
   • captcha_length - how long generated captcha is going to be
   • dataset_size - how large dataset is going to be generated
   • height - height of generated captcha
   • width - width of generated captcha

3. Generate annotations files using bin/captcha_annotate.py script. You can call it for example

   python bin/captcha_annotate.py --val_split=0.1 --test_split=0.1 --case_sensitive

   This will shuffle and split data into train/validation/test according to following parameters:

   • val_split - how large part of data is going to be used for validation, e.g. 0.1 means 10%
   • test_split - how large part of data is going to be used for testing
   • case_sensitive - switch denoting that labels that are created will be case sensitive
     • if such parameter is not passed, then for example if aBcD is in the image (and image is named accordingly), resulting label will be abcd
     • if it is passed, resulting label will be aBcD

   This script will create annotations.txt, annotations-train.txt, annotations-validation.txt and annotations-test.txt.

4. Run training script bin/train.py for example like this:

   python bin/train.py --available_chars="abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ" --captcha_length=6 

   Training script notably logs models after each checkpoint into logs/train.py-{START TIMESTAMP}-{parameters etc.} directory.