Multidimensional, complicated and dynamic outputs with dynamic object detection in keras

Question

I am trying to create a neural network in keras which receives a grayscale-image with a monocrome background color and handwritten digits, as well as symbols that are no digits,