This is a project only for SCUer, It uses a high accuracy model to recognize captcha and will input to input bar automatically.
I didn't upload the model because it's just too large to upload.
You could use CNNtrain.py to train your model. Just change the POS(in the beginning of CNNtrain file) to train the recognizer to recognize each pos and put the result together.
Since the scu zhjw only use 0 to y to create captcha, my model was designed to fit that. But if your captcha is generated by much more characters, just change the output units number of output layer and design your preprocessing of image.
Thanks my schoolmate for giving me datasets.