No you can't ... it will be slow ass as hell. What I suggest is the following approach
- Start with a VJ face detector. That reduces the search area.
- Using a mouth region detector would even be better, but might be hard to train.
- Use a classifier to define what mood is being expressed. Or apply a feature point detector to find mouth corners and define relations for moods.
Good luck!