How is Keras_ocr producing the detection boxes around the words?

Question

I am trying to figure out how keras\'s OCR builds a box around it\'s detections.

So far this is what I know:

On their GitHub (i.e. more specifically in keras_o