How can i identify the image of a hat on top of a face

edit

There are two possible approaches

1. Create a hat detector using the traincascade interface and detect hats instead of faces
2. Use a pretained face detector, run the detector and retrieve the detections, Then cut out an average hat region above the detection and run a hat/no-hat classifier on top of that region.

Haha and what is the scientific addition of this? :D