I'm partial to Dorkwood.

If you're going to use this approach, you need to have it also train on images from a large dataset like something from shutterstock or google imagesearch.