By what principle are the weights of a conventional neural network distributed when classifying an image?

Question

Python

I don’t understand a little how the weights are placed in the input layer of the neural network when classifying an image, because how can you determine the importance of a single pixel so that it all works also on test data?