All the needed files are inside the API_KERAS folder. In order to run the API locally just run the application_keras.py and use Postman to make a POST request.

If you want to download the container directly from my Docker Hub profile just run the following code:

docker pull dangape/api_keras

In order to test the container just run the image using the following command:

docker run dangape/api_keras

In order to test the API you can use postman or another request app to make a POST request. Notice that you can make a POST request with a base64 string image and with the key string, just like the image bellow.
Or you can make a POST request with a image file, just make sure to change the key name to 'file' in this case. Also remember to change the request code line inside the application_keras.py file. You can use this website to create a base64 string

You can find all the requirements to run the code in the requirements.txt file. But to make it easier I´ll list them below:

• numpy
• Flask
• opencv-python
• imageio
• Pillow
• imutils
• uwsgi
• tensorflow
• keras
• sklearn

This is a work in progress and currently the algorithm works just for simple CAPTCHAS, preferably with five letters.
Below you can see 4 examples of images that the algorithm can handle well.

• LETTER_LAB.py perform opencv tests on captcha files, with this file you can test process to see how they; will handle the given captcha.
• get_captcha.py perform web scrapping in some sites to download captcha files to train the model;
• label_data.py uses an online API to label downloaded data;
• build_training_data_letters.py split captcha into single letters and assign them to a folder with the matching letter name;
• train_model.py trains CNN model;
• test_accuracy.py tests the model accuracy;
• application_keras.py run the Flask API;
• processing_lab.py contains different models for processing different captchas;

• Use model1 for the first and second captcha type;
• Use model2 for the third captcha type;
• Use model3 for the fourth captcha type.

• model1: 87,5%
  • first captcha: 87,5%
  • second captcha: 69,5%
• model2: 45,00%
• model3: 99,00%