Recognition of numbers inside CAPTCHA using CNN neural network model. (using tensorflow library)
We have 101 different captchas and there are 10 different types of each captcha. each image has 5 digits, which is divided into 5 equal parts, and finally the model is trained on 5050 digit images.

So, we have trained the model on the numbers, not on the entire captcha image.

In the last layer of the model, softmax is used due to the existence of several classes for each data, and this gives us the probability of occurrence of each class.

To evaluate the model, sparse categorical crossentropy is used and a validation set is given to the model during training to calculate its loss and accuracy. Finally, the accuracy of the model was 99% on the training data and 96% on the test data.