The subsequent steps I would follow are
- After extracting the face, detect facial landmarks, for example by using dlib
- Then extract features around the chin regio's together with image patches
- Create a local descriptor, for example an LBP histogram
- Then do SVM or ANN classification between classes
