It depends on what you actually want!
- If you want an actual eye+glass detector, then crop all your training data to the eyes with glasses.
- If you want a face detector, that is able to detect face with and without glasses make a dataset with full faces, both with and without glasses. This will force the system to select features that are common between both classes.
- If you want to seperate between faces / glassfaces, then cascade classifiers are not the way to go. Then you will need to do some post processing filter after face detection to determine if a glass region is there or not. For that make a second classifier, for example an SVM on a cut out eye or eye+glass region as a binary decision.
