Click image to see original.

Requires Python 2, Python Imaging Library, Numpy.

The basic principle is to draw lines but not corners or dots. Lines, and corners and dots can be detected by the eigenvalues and determinant of the Hessian respectively. I also need to blur the image for various reasons, which I attempt to do in as scale-free a way as possible by using a long tailed convolution kernel (of course, you know I'm obsessed by these things), a t-distribution to be precise.