Image processing has been since many years at the forefront of computer science and the recent, exciting developments in deep learning applied to computer vision aroused the interest in the field even more. However, while the old style image processing pipelines were very well understood and relied on solid mathematical arguments, the more modern neural networks often seem like a black magic solution to a badly defined problem. In this talk, I will try to shed some light on the matter, comparing old algorithms and new approaches, and show with simple examples how to build and train a network to solve a computer vision task.