Well, in theory you could train on a set of examples of features, and you could have the classifier reject an example if it was too many standard deviations outside the mean for the features. But this is a fairly weak classifier.
What would be much more common would be to train on one-versus-everything-else -- for example to identify stop signs, you would train with several examples of stop signs, and you would also include examples of other traffic signs... and of dalmations, and of blueberries, and of rivers, and of cereal boxes, and of sunsets, and of spectrograms... and so on.
The classification operation would then either predict class 1 (stop signs) or class 2 (image is something else that is not a stop sign).
